A Survey of Open Weights Image Generation Models Running Locally in 2026

Summary. No open-weights image model sits within one tier of the closed frontier on either public preference leaderboard, and the models that do lead the open board are frequently the ones a business cannot legally deploy. This piece surveys the full stack a team evaluating local image generation in 2026 has to work through: the base model lineages and the licensing decisions that produced them, the fine-tune and merge ecosystem that determines which models are actually usable in practice, the instruction-editing tier that did not exist eighteen months ago, the hardware and rental economics, and the license text governing each model, since the license attached to a repository is routinely disconnected from the license actually governing the weights inside it. Every figure below carries a source, and the numbers that conflict across sources are reported as conflicts rather than smoothed into a single confident claim.

What local image generation covers

"Local image generation" is used to mean at least seven distinct things, and a roundup that only covers one of them misses most of what a team evaluating this space actually needs.

Licensing runs across every one of these layers, and it is the single most common reason a technically correct model choice turns out to be a commercially unusable one.

The measured gap between open and closed models

Two blind-preference arenas score image models on the same method chess ratings use for players: an identical prompt, two anonymous outputs, a human selects the stronger one, and an Elo score results from the aggregate of those choices. On Artificial Analysis's open-weights board, Ideogram 4.0 leads at 1,222 Elo [1]. On arena.ai's unrestricted board, spanning 5,919,946 votes across 77 models, OpenAI's gpt-image-2 leads at 1,381, followed by Microsoft's mai-image-2.6-preview at 1,336, xAI's grok-imagine-image-2.0 at 1,316, and Reve 2.1 at 1,302 [2]. The gap between the open leader and the closed leader, 148 to 177 Elo points depending on which arena is read, is the figure most coverage of this space leads with.

A second question changes the ranking that matters for a deployment decision: which of these models a team may legally place inside a product. Ideogram 4.0's weights carry a non-commercial license [3], so the ranking for a commercial deployment begins several places lower, with a 20-billion-parameter Apache 2.0 model, Qwen-Image-2512, at 1,172 Elo, running on a single 24 GB GPU [4]. That gap carries a direct cost when it is misjudged, and it is the check most published model comparisons skip before publishing a ranking.

Open weights hold a concrete lead in one measurable area: rendering text inside an image. Three separate labs report the same result. GLM-Image scores 0.9116 on CVTG-2K word accuracy against Seedream 4.5's 0.8990 [5]. LongCat-Image reaches 90.7 on a Chinese-word rendering benchmark against a next-best of 58.5 [6]. Ideogram 4.0 won a blind test of ten professional designers 47.9% of the time against Nano Banana 2's 30.0% [3]. All three figures are self-reported by their own labs, and three independent labs reporting the same advantage functions as a degree of corroboration across separate evaluations.

Open weights also hold structural advantages no benchmark captures: the marginal cost of an image is electricity, there is no content policy sitting between a prompt and an output, training a LoRA on private data is possible, structural conditioning such as ControlNet and pose input has no equivalent in a closed API, and source images that must not leave a machine never have to.

What the open field is competing against

The closed hosted tier, versions and pricing as of this writing:

A cost anchor for the local-versus-hosted argument: GPT Image 2 at high quality runs about $211 per 1,000 images, and Nano Banana Pro runs about $134 per 1,000 [7]. A local FLUX.2 [klein] 4B run costs roughly the electricity of a 1.2-second generation on an RTX 5090.

Licensing decisions behind the Stable Diffusion and FLUX lineages

Stable Diffusion 1.5 shipped in October 2022 as an 860-million-parameter UNet under CreativeML OpenRAIL-M, a license permissive on money and restrictive on conduct. It names no commercial threshold, and instead lists prohibited uses in an attachment that downstream users are contractually obligated to keep enforcing [8]. SDXL followed in July 2023 on the same family of terms and remains, three years later, the only Stability model with a living fine-tune ecosystem, for reasons the next section sets out with figures.

Stable Diffusion 3 changed the terms. The mid-2024 release replaced the OpenRAIL-era terms with the Stability AI Community License, which requires registration and terminates automatically once a company crosses a million dollars in annual revenue, calculated on total company revenue rather than revenue attributable to the model [9]. Several independent accounts describe the same sequence following the change: a licensing backlash, the departure of a meaningful share of the original Stable Diffusion research team, and the founding of Black Forest Labs, which shipped FLUX two months later [10, 11].

FLUX then reproduced the same underlying problem in a more confusing form. FLUX.1 [dev] is free to download, technically strong, and carries a license that forbids running the weights for any purpose connected to payment, direct or indirect, while separately stating that any image generated with it belongs to the person who generated it [12]. Those two clauses separate the verb from the noun. Operating the model for a commercial purpose is restricted. The images the model produces are not restricted from resale by Black Forest Labs. Black Forest Labs' own model card adds to the confusion by stating that outputs "can be used for personal, scientific, and commercial purposes" directly beside a license titled Non-Commercial, and the community reads the contradiction in both directions across the model's own GitHub discussions [12]. A designer who runs FLUX.1 [dev] on a personal machine and invoices a client for the resulting image is, on the license's own terms, almost certainly still inside the restricted use, since the operation of the model was tied to compensation. A publisher who later purchases that finished image from the designer never touches the license at all.

Black Forest Labs corrected this for one model. FLUX.2 [klein] 4B, released January 2026, is Apache 2.0, runs in 8.4 GB, generates in about 1.2 seconds on an RTX 5090, and performs text-to-image generation and reference-based editing from a single checkpoint with no fine-tuning required [13]. It ranks sixth among open editors on GEditBench v2 despite its size, and it is the most commercially usable local model of comparable quality available today. It took Black Forest Labs eighteen months to ship it.

The full FLUX lineage, with the licensing pattern visible across every release:

Chroma is the community's response to the licensing problem: a de-distillation of FLUX.1 [schnell] back into a trainable, CFG-capable base under permissive terms [13]. It matters more as a statement of intent than as a usage story, since it pulls only a few hundred downloads a month from the maintainers tracked in the next section.

The rest of the field: base models outside Stability and Black Forest Labs

Twenty-plus base models shipped from Chinese labs, American startups, NVIDIA, and academic groups between 2024 and 2026, and most of them never reach a general audience despite carrying real capability.

Qwen-Image, and the moment it closed. Alibaba released the 20B Qwen/Qwen-Image model, two editors, and a layered variant under Apache 2.0 through the end of 2025, culminating in Qwen-Image-2512, the last open base in the line, and Qwen-Image-Edit-2511, the last open editor [14]. Qwen-Image-2512 has 183 published adapters and 37 finetunes; the edit line pulls roughly 462,000 downloads a month at its peak [14]. Qwen-Image-2.0 and Qwen-Image-3.0 then arrived in 2026 with no weights, no model card, and no benchmark disclosures; a maintainer wrote plainly that there was no plan to open-source the newer line [15]. Every pipeline and adapter built on the Apache-licensed 20B model now has no upgrade path inside that family.

Z-Image, the adoption winner. Tongyi-MAI's Z-Image-Turbo is a 6-billion-parameter, Apache 2.0 model that fits comfortably within 16 GB VRAM and runs in 8 steps [16]. Its own paper describes 20-to-80-billion-parameter contemporaries as "impractical for inference, and fine-tuning on consumer-grade hardware," positioning 6 billion parameters as a deliberate rejection of scale for its own sake [16]. The adoption figures support the argument: 780,384 downloads a month, 846 community adapters, and 148 finetunes, against 13,000 downloads and 2 finetunes for the 80-billion-parameter HunyuanImage 3.0 [16, 17].

HiDream-O1-Image, the best quality per parameter. An 8-billion-parameter, MIT-licensed model using a pixel-level unified transformer with no external VAE and no separate text encoder, scoring Elo 1,174 for text-to-image and 1,191 for editing, ahead of models several times its size [18]. No official VRAM figure has ever been published for it, and the authors report no GEdit-Bench or ImgEdit score despite the model's ranking.

Ideogram 4.0, the open-weights quality leader. A 9.3-billion-parameter, fully single-stream flow-matching diffusion transformer using a Qwen3-VL-8B-Instruct text encoder and a structured JSON-only prompting interface, the design choice behind its typography lead [3]. Its weights are gated under a non-commercial agreement, with a self-serve commercial license starting at $300 a month for 10,000 to 100,000 images.

Cosmos3-Super-Text2Image, the commercially usable frontier. NVIDIA's 64-billion-parameter mixture-of-transformers model, licensed under OpenMDW-1.1, a genuinely commercial license, requiring "Built on NVIDIA Cosmos" attribution [19]. It needs four to eight H100s or a single B200, which places it outside any consumer setup, but it is the strongest open-weights checkpoint a business may legally sell output from.

HunyuanImage 3.0, the model most people cannot license. An 80-billion-parameter, 13-billion-activated mixture-of-experts model that tops the open editing board at Elo 1,222, requiring roughly eight 80 GB accelerators for its full instruct variant [17]. Its license explicitly excludes the European Union, the United Kingdom, and South Korea from the grant; use in those jurisdictions is described in the license text as "unlicensed and unauthorized" rather than merely restricted.

Anima, the cleanest provenance in the anime lineage. Built by CircleStone Labs with Comfy Org on NVIDIA Cosmos-Predict2-2B rather than descending from a leaked checkpoint, with continuous development through anima-turbo-v1.1 on 2026-08-24 [20]. Its license, CircleStone Labs Non-Commercial v1.2, expressly defines LoRAs and textual inversions as derivatives, so fine-tunes built on it inherit the same restriction as the base model.

A representative slice of the wider field, ranked by Artificial Analysis Elo where a score exists:

Two further releases are worth naming for how differently their labs treated the fine-tune ecosystem. Meituan's LongCat-Image ships a -Dev variant explicitly described as a mid-training checkpoint "suitable for fine-tuning," the most community-aware packaging decision of any lab surveyed here [21]. Krea 2 ships a paired RAW and Turbo checkpoint with a documented workflow: train LoRAs on the undistilled RAW model, then apply them to the distilled Turbo model for inference [22]. Nobody else in the field documents that pairing as the intended path.

Fine-tune adoption measured across base models

Community claims about which base model fine-tuning actually concentrates on vary by source. A controlled comparison settles the question with figures rather than opinion. One prolific maintainer, SG161222, publishes photoreal checkpoints on four different bases through the same channel to the same audience: SDXL, SD 3.5 Medium, a de-distilled Flux, and Chroma. Monthly downloads for each: SDXL's RealVisXL reaches 414,000; the SD 3.5 tune reaches 15; the Flux tune reaches 6; the Chroma tune reaches 269 [23]. The maintainer's skill, distribution channel, and audience hold constant across all four figures. The variable that moved by four orders of magnitude is the base model's license and mergeability, not its quality.

SDXL earned that gravity because checkpoint merging is an established technique with a decade of tooling behind it: weighted sums, block-weighted merges with roughly 25 independent per-block weights, LoRA extraction and baking, all running through a single portable checkpoint file under a license that permits redistributing the result [24]. Flux breaks each of those preconditions. Its guidance distillation resists further training, which is the reason Chroma exists as a de-distilled, trainable rebuild of FLUX.1 [schnell] under permissive terms. Its non-commercial license forbids the merge business model directly, since a merged checkpoint is a derivative of a non-commercial model. Its 22-gigabyte weights, shipped with separately packaged T5-XXL and CLIP-L encoders, break the one-file convention the merge economy depends on. Flux's adapter ecosystem is therefore built from LoRAs rather than merge families: small, inexpensive to train, requiring no redistribution of base weights, with the licensing question left to whoever downloads them.

New development in 2026 moved past Flux for a separate reason and concentrated on Z-Image, for the arithmetic reason already covered above: a 6-billion-parameter model fits consumer VRAM for training, not only for inference. One brand's history illustrates the full arc. Juggernaut, among the most downloaded photoreal checkpoints ever published, started on SDXL under OpenRAIL, moved to Flux and became API-only once Flux's licensing made open redistribution impractical, and has since moved again, to Z-Image, under CC BY-NC 4.0 [16]. Even the model that captured new development in 2026 is being fine-tuned into non-commercial derivatives. The licensing constraint moved to a different point in the stack rather than being resolved.

Named photoreal brands and their base lineage:

Legal exposure in the anime and illustration model lineage

Anime and illustration models form the largest fine-tune ecosystem in local image generation. Its two branches each carry a distinct legal exposure, independent of whether AI-generated output itself is copyrightable.

The family tree runs from SD 1.4/1.5 through a leaked NovelAI checkpoint into Anything, AbyssOrangeMix, and Counterfeit on one side, and from SDXL into Illustrious XL, NoobAI-XL, and most current illustration merges on the other, with a separate branch running from Lumina-Image 2.0 into Neta Lumina, and a third running from NVIDIA Cosmos-Predict2-2B into Anima.

The older branch traces to a leaked NovelAI checkpoint that entered the community through a 2022 security breach [25]. No license governs its use, because no license was ever granted for it.

The more capable modern branch carries a subtler obligation. Illustrious XL, and everything built on it including NoobAI-XL and most current illustration merges, is licensed under the Fair AI Public License 1.0-SD, a copyleft license whose text states that "to modify also means to perform any training on a model or to combine a model with another model," and that deploying a derivative as a network service triggers a duty to make the derived weights available to that service's users, at minimum as a download link [26]. A closed-weights commercial product built on an Illustrious-lineage merge does not fail on a technicality correctable later; the license makes the product non-compliant from the outset, under terms almost nobody using the ecosystem has actually read. The exit is Neta Lumina v1.0, an Apache 2.0 anime-capable model with no share-alike duty and no ancestry problem [26]. Anima carries the best documented training provenance of any anime-capable model and remains non-commercial without a negotiated agreement [20].

Beyond the license, three technical fault lines split the anime ecosystem internally: Danbooru and e621 tag-based prompting is a genuinely different interface from the natural-language prompting every 2026 base model expects, NoobAI-XL forked into separate v-prediction and epsilon-prediction branches that need different sampler settings, and every adapter is bound to a specific latent space, so leaving SDXL means leaving the largest adapter library in the ecosystem.

Instruction editing performance against the closed frontier

Instruction-based image editing, changing a described detail with no mask, did not exist as a local capability before June 2025 and has changed leaders twice since. FLUX.1 Kontext [dev], the model that established the category, now sits next to last among open editors on both major benchmarks [1, 27]. A team recommending it as a 2026 default is working from data fourteen months out of date.

Artificial Analysis's open-weights editing board, in order: HunyuanImage 3.0 Instruct at 1,222, HiDream-O1-Image at 1,191, FLUX.2 [klein] 9B at 1,166, Qwen-Image-Edit-2511 at 1,165, and FLUX.2 [dev] Turbo at 1,151 [1]. FLUX.1 Kontext [dev] sits at 1,016, below HiDream-E1.1's 977 and above only GLM-Image, Bagel, OmniGen V2, and the two Step1X-Edit predecessors.

The gap to the closed frontier here carries more detail than the raw text-to-image gap. GEditBench v2, a benchmark that scores instruction-following, output quality, and visual consistency separately, where consistency measures whether the model left the rest of the photo unchanged, finds GPT Image 1.5 leading instruction-following at 1,260 while scoring 846 on consistency, below every open model in its own top eight [27]. The best open instruction-following score, from Qwen-Image-Edit-2511, is 1,095, a 165-point gap against the closed leader, five times the 32-point gap in the models' overall scores [27]. Local editing models are, in aggregate, more conservative and more faithful to the parts of an image left untouched by the instruction, and weaker at resolving a genuinely difficult compositional instruction. That pattern is a basis for a deployment choice: local editing for retouching work where drift is the primary risk, and a hosted model for instructions that are themselves difficult to satisfy.

Individual editors worth naming beyond Qwen-Image-Edit-2511 and the FLUX.2 [klein] pair: Step1X-Edit-v1p2, which publishes the best-documented VRAM table in the field, running at 18 GB flat in FP8 plus offloading and reaching a KRIS-Bench score of 60.93 with thinking-and-reflection mode enabled [28]; FireRed-Image-Edit, strong on identity preservation and virtual try-on, though its multi-image mode beyond three references depends on a Gemini API call for region detection, which breaks the fully-local claim [29]; and ICEdit, a training-free approach that claims to run in 4 GB VRAM, the right choice when that ceiling is fixed and hard [30].

ImgEdit, a nine-operation benchmark, shows the field's shared weak point clearly: extracting an object with a clean alpha scores 1.70 to 2.35 for open models, against GPT-Image-1's already weak 2.90 [31]. Style transfer, by contrast, is saturated across the field at 4.19 to 4.93 and tells a buyer almost nothing.

Masks, identity, and restoration

Mask-based inpainting remains the right tool when a change needs to land in an exact location, which no instruction model reliably provides, since instruction models re-encode the whole image and can drift. FLUX.1 Fill [dev] conditions on both mask and image in the diffusion process across 384 input channels and handles inpainting and outpainting from a single checkpoint, though its license carries the same non-commercial restriction as FLUX.1 [dev] itself [32]. InstantX's Qwen-Image inpainting ControlNet, trained from scratch on 10 million images, handles object replacement, text modification, and outpainting [33].

Identity and reference preservation moved from adapters bolted onto a base model in 2024 to native capability inside the instruction editors in 2026. Qwen-Image-Edit-2511 was tuned specifically for multi-person group-photo consistency, and FLUX.2 [klein] advertises character, object, and style reference with no fine-tuning required [4, 13]. Adapters still matter for constrained VRAM, multi-identity scenes, and disentangling style from subject, where ByteDance's DreamO, UNO, USO, and UMO, all Apache 2.0 and built on FLUX or OmniGen2, have no in-model equivalent [34].

For upscaling and restoration, Real-ESRGAN remains the default: BSD-licensed, essentially zero hallucination by construction, and fast enough to run in a batch pipeline [35]. SeedVR2, Apache 2.0, is the field's preferred diffusion-based restorer when quality matters more than speed [36]. The rule that decides between them: if the upscaled output feeds LoRA training, fidelity matters more than perceptual quality, since hallucinated detail becomes a poisoned training sample, which favors regression and GAN upscalers like Real-ESRGAN over diffusion-based ones for that specific use.

Structural control and regional prompting

ControlNet coverage now varies sharply by base model. SDXL has the full original ControlNet zoo plus a consolidated Union model. FLUX.1-dev has Shakker-Labs' ControlNet-Union-Pro 2.0, trained from scratch on 20 million images, covering canny, soft edge, depth, and pose in one stackable checkpoint [37]. Qwen-Image has InstantX's Union ControlNet, though quality degrades under GGUF quantization. FLUX.2 does not yet have a ControlNet at all; its structural editing capability comes from reference latents built into the base model rather than a separate checkpoint, which is the direction the field is moving: consolidating what used to require a dozen separate ControlNet checkpoints into conditioning the base model already understands.

Regional prompting has three tiers of sophistication: latent-space compositing, which samples masked regions independently and blends them but shows seams; Attention Couple, which computes cross-attention separately per prompt region and is the more capable modern default; and Flux-specific region-attention tools that apply masks to the T5 text embeddings only. Krita AI Diffusion has shipped first-class regional prompting since v1.18.0 and extended it to Anima in v1.52.0, which is the least painful way to use the technique without wiring a custom graph [38].

Runtime, front-ends, and their maintenance status

The front-end a team picks determines how quickly a new model becomes usable, and the field has split sharply between actively maintained and abandoned projects.

Comfy Cloud, the hosted counterpart to ComfyUI, bills per second of active GPU time on RTX PRO 6000 Blackwell hardware with 900-plus models preinstalled, running from a free tier of five runs up to team pricing of $200 to $2,500 a month; the local application itself remains free with no feature gates [40]. AMD ROCm support became genuinely usable in 2026, with native integration into ComfyUI from Comfy Desktop v0.7.0 on ROCm 7.1.1, though Nunchaku's quantization format remains CUDA-only, leaving SD.Next's SDNQ as the cross-platform substitute [41].

Quantization, attention kernels, and distillation

Four quantization formats matter in 2026: FP8, a weight-only cast that is roughly half the size of BF16 with neutral to slightly faster inference; NF4, a 4-bit format giving 1.3 to 4 times speedup on 6 to 8 GB cards depending on the driver stack; GGUF, llama.cpp's block quantization adapted for diffusion transformers, which shrinks memory but does not by itself speed up weight-only inference; and SVDQuant, marketed as Nunchaku, the only format on this list that delivers a genuine speedup rather than only a size reduction, by quantizing both weights and activations to 4 bits and absorbing outliers into a separate low-rank branch [42].

The ICLR 2025 paper behind SVDQuant measured FLUX.1-dev at 25 steps on a 16 GB laptop RTX 4090: BF16 takes 111.7 seconds, naive 4-bit quantization drops to 12.5 seconds but with visibly worse output, and SVDQuant reaches 12.9 seconds while holding the lowest perceptual-distance score of the three [42]. Much of that speedup comes from eliminating CPU offloading entirely. The practical guidance from the format's own documentation: use INT4 on pre-Blackwell cards and NVFP4 on 50-series cards, since NVFP4 on a 5090 reaches roughly three times the speed of BF16 with better quality than INT4 [43].

For attention kernels, SageAttention claims two to five times the speed of FlashAttention on the same hardware, though the gain is inconsistent and sometimes negative depending on driver version and platform, and it fails silently rather than with an error message when a required compiler dependency is missing, occasionally masking a 20-times slowdown as normal behavior [44]. Attention itself accounts for 40 to 70 percent of total generation time in modern diffusion transformers, which is why these kernels matter as much as they do.

Distillation, the technique that reduces a 20-to-50-step model to 1-to-8 steps, has changed shape rather than disappeared. In 2024 it arrived as a bolt-on LoRA over an existing checkpoint. In 2026 it arrives pre-baked into the model itself: Z-Image-Turbo, Krea 2 Turbo, FLUX schnell and klein, and ERNIE Turbo all ship their distilled variant as the primary release rather than an add-on [16, 22]. The trade-off curve is consistent across labs: one step loses composition coherence and text fidelity, four steps is the practical production floor, and eight steps is close to indistinguishable from twenty to twenty-eight for most prompts.

Hardware requirements and the economics of renting versus buying

VRAM, not raw compute, is the constraint that determines what runs. FLUX.2 [klein] 4B fits an 8 GB card. A 20-billion-parameter model such as Qwen-Image requires 24 GB, less with GGUF quantization. HunyuanImage 3.0's 80-billion-parameter mixture-of-experts model requires eight 80 GB accelerators for its full instruct variant, which places it outside the range of any home setup [17].

VRAM tiers and what genuinely runs at each:

The clearest illustration of why VRAM headroom does not scale linearly: FLUX.2-dev at 8-bit quantization is roughly 30 GB of weights, and with BF16 activations it overshoots a single 32 GB card by about 1 GB per forward pass. On Windows, that overshoot pages through unified memory and throughput collapses from 14.4 seconds per iteration to 277 seconds per iteration within a few hundred steps; on Linux the same configuration simply fails to allocate [45]. A shortfall of roughly one gigabyte produces a nineteen-times slowdown.

No new consumer GPU generation has shipped as of this writing. RTX 50-series Blackwell cards launched in 2025 and remain current, with street prices that now track VRAM capacity rather than performance tier. Retail trackers disagree with each other by several hundred dollars within the same month, so the accurate way to state a price is as a range: a 16 GB RTX 5060 Ti runs $560 to $805 depending on the tracker, and a 32 GB RTX 5090 runs $4,288 to $4,830 [46, 47]. A used RTX 3090, still the value pick for 24 GB, runs $1,095 to $1,181 [46].

Renting is the stronger arithmetic for most workloads that are not continuous. Renting an RTX 3090 on Vast.ai at roughly $0.069 an hour breaks even against the used purchase price at 15,870 GPU-hours, about twenty-one years at two hours of generation a day [48]. Against RunPod's higher fixed rate of $0.50 an hour, the same card breaks even at roughly three months of continuous, round-the-clock use [48]. Purchasing is justified by latency, offline access, data privacy, and the fact that the card serves other purposes. On cost alone, purchasing is justified only at near-continuous utilization.

Apple Silicon removes the hard VRAM cliff through unified memory. A 64 GB Mac loads FLUX.1-dev at full FP16 with no quantization at all, which no consumer NVIDIA card can do. The cost is throughput: Apple's MPS backend runs two to four times slower than CUDA for the same operation, and generic PyTorch on Mac runs roughly three times slower than Draw Things' hand-tuned Metal path, which is why the software chosen on a Mac matters as much as the hardware [49].

LoRA training

AI-Toolkit is the 2026 default LoRA trainer, MIT-licensed, with the widest model coverage in the field: FLUX.1, FLUX.2 including klein, Qwen-Image, Z-Image and its variants, HiDream, Anima, and Krea 2 [50]. Kohya SS remains the most documented option and has the best low-VRAM story through its fused backward pass. OneTrainer has the strongest aggressive gradient-checkpointing implementation for 8 GB setups.

Training VRAM by base model, drawn from secondary sources since AI-Toolkit's own documentation publishes no per-model table:

A useful anchor: getting a Qwen-Image LoRA training on a 24 GB card required adding text-embedding caching so the text encoder could be unloaded during the training step; a prior attempt on a 32 GB card had needed 6-bit quantization just to fit [51].

Dataset sizing follows a simple rule of thumb: 20 to 50 images for a character or face, 50 to 200 or more for a style or object, roughly 100 training steps per image, and every image needs a caption. Cost, calculated from the rental rates above: a character or style LoRA on a rented RTX 4090 or 5090 costs under three dollars of GPU time and takes an afternoon, with the real cost sitting in dataset curation and captioning rather than compute, a step typically repeated several times before the result is acceptable [48].

Open models with licensing terms worth reading before deployment

Six releases carry more practical weight than their public profile suggests, and each comes with a licensing detail worth confirming before building on it.

A deployment decision table by license and hardware tier

The rule that governs all of the above is the one detail benchmark tables never carry: the governing document is the license file itself, independent of the repository tag or the bullet points on the model card. One model in this survey is tagged apache-2.0 in its own metadata while its LICENSE.txt states a restrictive, non-commercial NVIDIA license underneath [52]. That discrepancy recurs across the ecosystem more often than any other, and it is the detail every roundup built around a benchmark table omits.

Numbers worth double-checking before relying on them

A survey this wide accumulates disagreements between sources, and the honest way to handle them is to name them rather than average them away.

- HunyuanImage 3.0's parameter count is stated as "over 80 billion, with 13 billion activated" in its own paper and as "83B params" on its Hugging Face sidebar. The paper's figure is the primary source. - Qwen-Image-2.0's specifications circulate on several enthusiast sites with a 7B parameter count and a 16 GB VRAM requirement, but no repository has ever been published under that name. Treat the weights as unreleased until a repository appears. - HiDream-I1's VRAM requirement is reported anywhere from 12 GB to 80 GB depending on the source, with no single figure confirmed by the model's own documentation. - The RTX 50 Super refresh's cancellation is leak-sourced and contradicted by a later leak claiming the series is still planned. State only the verified part: no new consumer generation had shipped as of this writing. - No primary 2026 TensorRT benchmark for diffusion models exists in any source checked for this piece. Any TensorRT speedup figure circulating online should be treated as unverified. - SANA 1.5's own repository tags itself apache-2.0 in metadata while its LICENSE.txt is the restrictive NVIDIA license discussed above, inside the same repository.

References

1. Artificial Analysis. "Image Arena, Text-to-Image and Editing, Open Weights." artificialanalysis.ai, accessed August 25, 2026.
2. arena.ai. "Text-to-Image Leaderboard." arena.ai/leaderboard, snapshot dated August 10, 2026.
3. Ideogram AI. "Ideogram 4.0" model card and blog post. github.com/ideogram-oss/ideogram-4 and ideogram.ai/blog/ideogram-4.0, June 3, 2026.
4. Qwen Team, Alibaba. "Qwen-Image-2512" and "Qwen-Image-Edit-2511." huggingface.co/Qwen, December 2025.
5. Z.ai. "GLM-Image." github.com/zai-org/GLM-Image, January 2026.
6. Meituan. "LongCat-Image." arXiv:2512.07584, December 2025.
7. Artificial Analysis. "Image Editing Leaderboard, Pricing." artificialanalysis.ai/image/leaderboard/editing, accessed August 25, 2026.
8. Stability AI. "CreativeML Open RAIL++-M License." huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md, July 26, 2023.
9. Stability AI. "Stability AI Community License Agreement." stability.ai/community-license-agreement, accessed 2026.
10. arte.itlibra.com. "What Is Stable Diffusion." May 28, 2026.
11. digital-humans.org. "Stable Diffusion, the Open Image Model Family." June 2, 2026.
12. Black Forest Labs. "FLUX.1 [dev] Non-Commercial License," v1.1.1 and v2.0. github.com/black-forest-labs/flux and bfl.ai/legal/non-commercial-license-terms.
13. Black Forest Labs. "FLUX.2 [klein]." bfl.ai/models/flux-2-klein and github.com/black-forest-labs/flux2, January 15, 2026.
14. Qwen Team, Alibaba. "Qwen-Image" repository and issue tracker. github.com/QwenLM/Qwen-Image, 2025 to 2026.
15. QwenLM. GitHub issue #264. github.com/QwenLM/Qwen-Image/issues/264, 2026.
16. Tongyi-MAI, Alibaba. "Z-Image" technical report and model card. arXiv:2511.22699, huggingface.co/Tongyi-MAI/Z-Image-Turbo, November 27, 2025.
17. Tencent. "HunyuanImage-3.0." huggingface.co/tencent/HunyuanImage-3.0 and arXiv:2509.23951, September 28, 2025.
18. HiDream.ai. "HiDream-O1-Image." github.com/HiDream-ai/HiDream-O1-Image, May 8, 2026.
19. NVIDIA. "Cosmos3-Super-Text2Image." huggingface.co/nvidia/Cosmos3-Super-Text2Image, May 31, 2026.
20. CircleStone Labs and Comfy Org. "Anima" and license text. huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md, 2026.
21. Meituan. "LongCat-Image" repository. github.com/meituan-longcat/LongCat-Image, December 2025.
22. Krea AI. "Krea 2" repository. github.com/krea-ai/krea-2, June 2026.
23. SG161222. Checkpoint download statistics across RealVisXL, SD 3.5, Flux, and Chroma tunes. huggingface.co/SG161222, accessed August 25, 2026.
24. SuperMerger and the A1111/Forge merge tooling ecosystem, accessed 2026.
25. Community accounts of the January 2022 NovelAI model leak, as referenced across Stable Diffusion anime fine-tune histories, 2026.
26. Free Development Project. "Fair AI Public License 1.0-SD." freedevproject.org/faipl-1.0-sd, and Park et al., "Illustrious: An Open Advanced Illustration Model," arXiv:2409.19946.
27. StepFun, NTU, SEU. "GEditBench v2." zhangqijiang07.github.io/gedit2_web and arXiv:2603.28547, March 2026.
28. StepFun. "Step1X-Edit." github.com/stepfun-ai/Step1X-Edit and arXiv:2511.22625, November 2025.
29. Xiaohongshu (FireRed Team). "FireRed-Image-Edit." github.com/FireRedTeam/FireRed-Image-Edit and arXiv:2602.13344, February 2026.
30. "ICEdit / In-Context Edit." arXiv:2504.20690, NeurIPS 2025.
31. "ImgEdit" benchmark. arXiv:2505.20275, NeurIPS 2025.
32. Black Forest Labs. "FLUX.1 Fill [dev]." huggingface.co/black-forest-labs/FLUX.1-Fill-dev, November 2024.
33. InstantX. "Qwen-Image-ControlNet-Inpainting." huggingface.co/InstantX, 2026.
34. ByteDance. "DreamO," "UNO," "USO," "UMO" repositories. github.com/bytedance, 2025 to 2026.
35. "Real-ESRGAN." github.com/xinntao/Real-ESRGAN, ongoing maintenance.
36. ByteDance Seed and NTU. "SeedVR2." arXiv:2506.05301, ICLR 2026.
37. Shakker Labs. "FLUX.1-dev-ControlNet-Union-Pro-2.0." huggingface.co/dadufo/FLUX.1-dev-ControlNet-Union-Pro-2.0, 2026.
38. Krita AI Diffusion. github.com/Acly/krita-ai-diffusion, release history through v1.53.0, August 2026.
39. ComfyUI. github.com/comfyanonymous/ComfyUI, release history through v0.33.1, August 2026.
40. Comfy Org. "Comfy Cloud Pricing." comfy.org/cloud/pricing, accessed August 25, 2026.
41. Comfy Org and AMD. "Official AMD ROCm Support Arrives." blog.comfy.org, 2026.
42. "SVDQuant" (Nunchaku). ICLR 2025 proceedings, hanlab.mit.edu/projects/svdquant.
43. Nunchaku documentation. nunchaku.tech/docs, accessed August 25, 2026.
44. thu-ml. "SageAttention." github.com/thu-ml/SageAttention, accessed August 25, 2026.
45. ostris/ai-toolkit, GitHub pull request #829, discussion of FLUX.2-dev memory overshoot on 32 GB cards, 2026.
46. BuySellRam. "Nvidia Consumer GPU Price Report, August 2026." buysellram.com, August 7, 2026.
47. TechPowerUp and capitalandcompute.net GPU price tracking, August 2026.
48. Vast.ai and RunPod public hourly GPU rental pricing, accessed August 25, 2026.
49. Draw Things and MLX benchmark comparisons on Apple Silicon, insiderllm.com, 2026.
50. ostris. "AI-Toolkit." github.com/ostris/ai-toolkit, accessed August 25, 2026.
51. Ostris AI. LoRA training walkthrough on Qwen-Image, video documentation, 2026.
52. NVIDIA. "SANA 1.5" model repository, metadata tag versus LICENSE.txt discrepancy. huggingface.co/Efficient-Large-Model/SANA1.5_1.6B_1024px, accessed August 25, 2026.

Last updated: August 25, 2026