There is no single best ComfyUI model; pick by job and license. For realistic people you can post commercially, start with Z-Image Turbo or FLUX.2 [klein] 4B (both Apache 2.0, both on consumer GPUs) and train your character LoRA on Z-Image base or klein 4B base. Krea 2 is the newest open option, with commercial use capped at $1M in annual revenue. Qwen-Image handles text inside images on a 24 GB-class GPU. FLUX.1 [dev], FLUX.2 [dev], Qwen-Image-2.1 and Ideogram 4 are non-commercial. For anime, SDXL fine-tunes such as Illustrious-XL keep the largest LoRA ecosystem. For local video, Wan 2.2 (Apache 2.0) is the default, with LTX-2.5 when you need audio.
The best ComfyUI models for most people in October 2026 are Z-Image Turbo and FLUX.2 [klein] 4B: both are Apache 2.0, both run on consumer GPUs, and both come with an undistilled sibling you can train a character LoRA on. Everything else on this page is a better choice only for a specific job, such as text inside images (Qwen-Image), anime (SDXL fine-tunes) or video (Wan 2.2), or carries a license that limits what you can do with it.
Checked on October 11, 2026 against official model cards and license files on Hugging Face, vendor repositories and announcements, and docs.comfy.org. Nothing here is our own benchmark: we have not scored these models' output side by side, so the order follows documented properties (license, VRAM, fine-tuning support), not a quality ranking. This update adds Krea 2, Ideogram 4, Z-Image base, Qwen-Image-2.1 Turbo and the status of FLUX 3 to the October 1 version.
This page is the model-choice chapter of our AI image generation for influencers guide. If you have not installed anything yet, what ComfyUI is comes first.
Best ComfyUI models by use case
- Realistic people you can post commercially: Z-Image Turbo for speed; Z-Image (base) or FLUX.2 [klein] 4B base to train a character LoRA. All Apache 2.0.
- Fast images you can sell: Z-Image Turbo (6B, Apache 2.0) or FLUX.2 [klein] 4B (Apache 2.0).
- Text inside images, posters and editing: Qwen-Image and Qwen-Image-Edit-2511 (20B, Apache 2.0).
- The newest open aesthetic model: Krea 2 Turbo (12B), if your annual revenue is under its $1M license cap.
- The Flux look for personal projects: FLUX.1 [dev] or FLUX.1 Krea [dev]; FLUX.2 [dev] if you have the hardware. All three are non-commercial models.
- Anime and stylized characters: SDXL fine-tunes such as Illustrious-XL, or Pony V7. Check each checkpoint's license.
- Low VRAM and the biggest LoRA library: SDXL 1.0 and its fine-tunes.
- Open video: Wan 2.2 (Apache 2.0); LTX-2.5 for video with synchronized audio; HunyuanVideo 1.5 as a smaller option.
Every model in the tables below runs in ComfyUI without custom nodes. VRAM figures are the vendor's or Comfy's own statements, not our benchmarks.
- 01Job
People, text, anime, edits or video
- 02License
May you post or sell the result?
- 03VRAM
Does a published figure fit your card?
- 04LoRA base
Is there an undistilled variant to train on?
- 05Template
Load the official workflow
Best ComfyUI checkpoint for realistic people
For realistic people, start with Z-Image or FLUX.2 [klein] 4B and move down the list only for a reason. The list is ordered the way a creator has to decide: license first, then whether the model fits your GPU, then whether you can train your character into it. It is not ordered by image quality, which we have not measured, and you should be wary of any ranking that names a winner without showing its test.
| Pick | Why it is on the list | VRAM guidance (official) | License and commercial use |
|---|---|---|---|
| 1. Z-Image Turbo | Its card says it "excels in photorealistic image generation"; 8 function evaluations, so drafts are fast | "Fits comfortably within 16G VRAM consumer devices" | Apache 2.0: yes |
| 2. Z-Image (base) | Undistilled; its card calls it "a good base for LoRA training"; CFG and negative prompts work | No official figure; same 6B size as Turbo | Apache 2.0: yes |
| 3. FLUX.2 [klein] 4B | Generation and multi-reference editing in one model; a 4-step distilled variant and a base variant for LoRA training | About 8 GB (BFL GitHub) or 13 GB (card); Comfy measured 8.4 GB distilled, 9.2 GB base | Apache 2.0: yes |
| 4. Krea 2 Turbo, with LoRAs trained on Krea 2 Raw | Released June 22, 2026; Comfy says its range runs "from photographic portraits to anime"; 8 steps | No official figure; Comfy's Turbo files are 12.2 GB (fp8) and 7.2 GB (nvfp4) | Krea 2 Community License: commercial use only under $1M annual revenue |
| 5. SDXL fine-tunes: Juggernaut XL v9, RealVisXL V5.0 | Both cards state a photorealism aim; the deepest library of LoRAs and ControlNets | SDXL: "consumer GPUs with 8GB VRAM" | Juggernaut XL v9: CreativeML Open RAIL-M, no paid API hosting without a license. RealVisXL V5.0: OpenRAIL++ |
| 6. Qwen-Image (20B) | The pick when the picture must contain readable text; heavy | Comfy's fp8 test used 86% of a 24 GB RTX 4090D | Apache 2.0: yes |
| 7. FLUX.1 [dev] and FLUX.1 Krea [dev] | The original Flux look, with a large library of community LoRAs | Original file about 23 GB; Comfy suggests about 24 GB for Krea's original weights | Non-commercial model license: read the Flux license section |
| 8. Qwen-Image-2.1 and 2.1 Turbo | 7B, generation and editing in one model, 2K output; Turbo (October 9, 2026) runs in 8 steps | No official figure; Comfy's bf16 file is 13.3 GB | Qwen Research License: research or evaluation only |
- Picks 1 to 3 are the default. Each family pairs a fast distilled model with an undistilled one. BFL's repository says to use the klein base variant "for fine-tuning, LoRA training, and maximum flexibility," and the Z-Image card marks Turbo as not fine-tunable, so train on the base and draft on the fast one.
- Pick 4 works the same way. Comfy's Krea 2 docs put it in one line: "train LoRAs on RAW, run inference on Turbo." The license is the catch, covered below.
- Pick 5 is for 8 GB cards and existing SDXL libraries. Read the version's license, because the Juggernaut line changes: v9 is CreativeML Open RAIL-M, Juggernaut XI is CC BY-NC-ND 4.0, and the 2026 Juggernaut Z (a Z-Image fine-tune) is CC BY-NC 4.0.
- Picks 6 to 8 are for a specific need or for personal work. Qwen-Image earns its VRAM when text matters; the last two are restricted by license, not by quality.
Source: Model cards and license files, checked October 2026
Once the base is chosen, the persona itself is a workflow problem: our ComfyUI AI influencer workflow shows how the base model, the character LoRA, the face detailer and the upscaler connect.
Image models compared
| Model | Released | Size | License | Commercial use | Official VRAM guidance |
|---|---|---|---|---|---|
| Z-Image Turbo | Nov 2025 | 6B | Apache 2.0 | Yes | "Fits comfortably within 16G VRAM consumer devices" (Tongyi-MAI) |
| Z-Image (base) | Jan 2026 | 6B | Apache 2.0 | Yes | No official figure |
| FLUX.2 [klein] 4B | Jan 2026 | 4B | Apache 2.0 | Yes | About 8 GB (BFL GitHub) or 13 GB (model card); Comfy measured 8.4 GB for the distilled version |
| FLUX.2 [klein] 9B | Jan 2026 | 9B | FLUX Non-Commercial License | No, without a BFL license | No official figure |
| FLUX.2 [dev] | Nov 2025 | 32B | FLUX Non-Commercial License | No, without a BFL license | Full model needs an "H100-equivalent GPU"; quantized builds target consumer GPUs |
| FLUX.1 [dev], Krea [dev], Kontext [dev] | Aug 2024 to Jul 2025 | 12B | FLUX.1 [dev] Non-Commercial License | Restricted; see the license section | Original FLUX.1 file is about 23 GB; Comfy suggests Krea's original weights with about 24 GB |
| FLUX.1 [schnell] | Aug 2024 | 12B | Apache 2.0 | Yes | No official figure |
| Krea 2 Raw and Turbo | Jun 2026 | 12B | Krea 2 Community License | Only below $1M annual revenue; enterprise license above | No official figure; Comfy's Turbo files are 12.2 GB (fp8) and 7.2 GB (nvfp4) |
| Qwen-Image, Qwen-Image-Edit-2511 | Aug 2025, Dec 2025 | 20B | Apache 2.0 | Yes | No official figure; Comfy's fp8 test used 86% of a 24 GB RTX 4090D |
| Qwen-Image-2.1 and 2.1 Turbo | Sep 2026, Oct 2026 | 7B generator | Qwen Research License | No (research or evaluation only) | No official figure |
| Ideogram 4 | Jun 2026 | 9.3B | Ideogram 4 Non-Commercial | No | No official figure |
| HiDream-I1 | Apr 2025 | 17B | MIT (bundled Llama 3.1 encoder has its own license) | Yes | fp8 files need more than 16 GB, full files more than 27 GB (Comfy docs) |
| Chroma1-HD | Aug 2025 | 8.9B | Apache 2.0 | Yes | No official figure |
| Stable Diffusion 3.5 Large / Medium | Oct 2024 | 8.1B / 2.5B | Stability AI Community License | Free below $1M annual revenue; register with Stability first | Medium needs 9.9 GB, excluding text encoders |
| SDXL 1.0 | Jul 2023 | 3.5B base | CreativeML Open RAIL++-M | Yes, within its use restrictions | "Should work effectively on consumer GPUs with 8GB VRAM" |
Sources: model cards for Z-Image Turbo, Z-Image, FLUX.2 [klein] 4B, [klein] 9B, FLUX.1 [schnell], Krea 2 Turbo, Qwen-Image-Edit-2511, Qwen-Image-2.1, Ideogram 4, HiDream-I1, Chroma1-HD and SDXL; BFL's FLUX.2 repository; the Krea 2 Community License; Stability's SD 3.5 and SDXL announcements and Community License; and Comfy's tutorials for FLUX.1, FLUX.2 [klein], Z-Image, Krea 2, Qwen-Image, Qwen-Image-2.1 and HiDream-I1. File sizes are from the Comfy-Org repositories.
Z-Image Turbo and FLUX.2 [klein] 4B: the practical defaults
Both are small, fast and Apache 2.0, which makes them the safest starting points for creators who plan to sell or publish commercially. Z-Image Turbo is a distilled 6B model that needs only 8 function evaluations per image and also renders English and Chinese text. FLUX.2 [klein] 4B combines generation and editing, including multi-reference edits, in one model. The larger [klein] 9B is non-commercial, so check which file you download. Z-Image (base), released in January 2026, is the undistilled version: its card recommends 28 to 50 steps at a guidance scale of 3 to 5 and describes it as a base for LoRA training.
Krea 2: the newest open model, with a revenue cap
Krea released Krea 2 as open weights on June 22, 2026: a 12B model in two checkpoints, Raw for training and an 8-step Turbo for generation. ComfyUI supports it natively. Its Community License permits commercial use only if your company-wide annual revenue is under $1,000,000, says you own the outputs you generate, and requires deployers to run content filtering. Above the cap you need an enterprise license from Krea.
Qwen-Image: text and editing
Qwen-Image is a 20B model known for rendering text inside images, and its Edit versions (2509, 2511) handle instruction-based edits. It is heavier than Z-Image or [klein] 4B. The newer Qwen-Image-2.1 is a 7B model that generates and edits at up to 2K with transparent backgrounds, and an 8-step Turbo checkpoint followed on October 9, 2026, but the Qwen Research License on both allows research or evaluation only. See our Qwen Image Edit guide for the ComfyUI workflow.
FLUX.1 and FLUX.2 [dev]: quality under a non-commercial license
FLUX.1 [dev] and its Krea and Kontext variants remain popular for photoreal images and instruction editing, and FLUX.2 [dev] is Black Forest Labs' largest open-weight image model at 32B. All of them carry non-commercial model licenses, covered below. Our Flux Kontext guide covers the editing workflow, and Flux consistent character covers keeping one face across images.
- FLUX.1 [dev], Krea [dev], Kontext [dev]: non-commercial model license
- FLUX.2 [dev] and [klein] 9B: FLUX Non-Commercial License
- Qwen-Image-2.1 and 2.1 Turbo: research or evaluation only
- Ideogram 4: non-commercial weights
- Krea 2: commercial use only under $1M annual revenue
- SD 3.5: free below $1M annual revenue
- NoobAI-XL: no commercialization
- Z-Image and Z-Image Turbo: Apache 2.0
- FLUX.2 [klein] 4B: Apache 2.0
- FLUX.1 [schnell]: Apache 2.0
- Qwen-Image and Qwen-Image-Edit: Apache 2.0
- Chroma1-HD: Apache 2.0
- HiDream-I1: MIT
- Wan 2.2 (video): Apache 2.0
Source: License tags and files on each model page, checked October 11, 2026
Best Flux models for ComfyUI
The best Flux model to run in ComfyUI today is FLUX.2 [klein] 4B, because it is the only current Flux model that is both Apache 2.0 and sized for consumer GPUs. The rest of the family is either larger and non-commercial or not available as weights at all.
| Flux model | What it is | License | Runs locally in ComfyUI? |
|---|---|---|---|
| FLUX.2 [klein] 4B | Fast generation and multi-reference editing; distilled (4 steps) and base variants | Apache 2.0 | Yes. The practical default |
| FLUX.2 [klein] 9B and 9B KV | The larger klein; BFL says 9B KV is faster for editing at equal quality | FLUX Non-Commercial License | Yes, after accepting the agreement on Hugging Face |
| FLUX.2 [dev] | 32B generation with single- and multi-reference editing | FLUX Non-Commercial License | Yes, with quantized files on consumer cards |
| FLUX.1 [dev], Krea [dev], Kontext [dev] | The 12B originals: photoreal images and instruction editing | FLUX.1 [dev] Non-Commercial License | Yes |
| FLUX.1 [schnell] | The 12B few-step model | Apache 2.0 | Yes |
| FLUX 3 Image | Announced July 23, 2026 | No open image weights yet; FLUX 3 Dev is on BFL's roadmap | Only as a paid Partner Node, added in ComfyUI v0.39.0 |
FLUX 3 is the name to be careful with. Black Forest Labs announced it on July 23, 2026 and listed an open-weight "FLUX 3 Dev" among capabilities it would release over the following weeks and months. As of October 11, 2026 the only FLUX 3 weights on BFL's Hugging Face page are the FLUX 3 Action robotics models. ComfyUI's v0.39.0 release (October 5) added FLUX 3 Image as a Partner Node, which means a paid API call, not a local model.
FLUX licenses, read carefully
The FLUX.1 [dev] Non-Commercial License v1.1.1 grants use of the model "solely for your Non-Commercial Purposes" and says use "for revenue-generating activity" or "in direct interactions with or that has impact on end users" is not one. Its Outputs clause says "You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein," but its Restrictions section bars using the model "or any data produced by" it "for any commercial or production purposes." The FLUX.2 [dev] and [klein] 9B license follows the same pattern, and the FLUX.2 [dev] model card says generated outputs "can be used for personal, scientific, and commercial purposes" as that license describes.
In practice: personal experiments and testing are fine; client work, a paid app or images for a monetized account are commercial use of the model and need a license from Black Forest Labs. If you want to avoid the question, use FLUX.1 [schnell], FLUX.2 [klein] 4B or another Apache 2.0 model. This is a summary of the license text, not legal advice.
Anime and stylized models
For anime and illustration, SDXL fine-tunes still lead because most anime LoRAs, ControlNets and style packs target them. They run on the same modest hardware as SDXL.
- Illustrious-XL: an SDXL-based anime model from OnomaAI. Its latest open weights, v2.0 (April 2025), are tagged CreativeML Open RAIL-M on Hugging Face.
- NoobAI-XL: another SDXL anime fine-tune. Its model card says: "We prohibit any form of commercialization, including but not limited to monetization or commercial use of the model, derivative models, or model-generated products."
- Pony V7: built on AuraFlow rather than SDXL. Its card summarizes the Pony License: you can use the model and outputs commercially unless you provide an inference service or app, have a company with over $1M in revenue, or use it in professional video production. The older Pony Diffusion V6 XL is not on PurpleSmartAI's Hugging Face page, so read the license wherever you download it.
Many anime checkpoints on community sites are further fine-tunes of these, and each uploader sets its own permissions. Read the license on the exact file you download.
Video models compared
| Model | Released | Size | License | Official VRAM guidance |
|---|---|---|---|---|
| Wan 2.2 TI2V-5B | Jul 2025 | 5B | Apache 2.0 | Comfy: "should fit well on 8GB vram with the ComfyUI native offloading"; Wan's own script needs 24 GB |
| Wan 2.2 T2V and I2V A14B | Jul 2025 | 27B total, 14B active | Apache 2.0 | Wan's scripts: at least 80 GB. Comfy tested its Wan 2.2 Fun Control workflow on a 24 GB RTX 4090D |
| Wan 2.2 Animate and S2V | Aug to Sep 2025 | 14B | Apache 2.0 | No official figure |
| MiniMax H3 | Jul 2026 | 33B | MiniMax H3 Community License; commercial use of locally generated outputs needs a MiniMax license | No official figure; Comfy's templates ship int8 checkpoints |
| HunyuanVideo 1.5 | Nov 2025 | 8.3B | Tencent Hunyuan Community License (excludes the EU, UK and South Korea) | "Minimum GPU Memory: 14 GB (with model offloading enabled)" |
| LTX-2, LTX-2.3 and LTX-2.5 | Jan to Aug 2026 | 19B (LTX-2), 22B (2.3 and 2.5) | Lightricks community licenses: free below $10M annual revenue, paid license above | Lightricks: full dev model 80 GB+; distilled and fp8 variants run on 32 GB and on as little as 12 GB |
Sources: the Wan 2.2 repository, Comfy's Wan 2.2 and Wan 2.2 Fun Control tutorials, the MiniMax H3 card and Comfy's H3 guide, the HunyuanVideo 1.5 card, the LTX-2.5 card and license, Comfy's LTX-2.5 tutorial and Lightricks' model information page.
Wan 2.2's repository adds: "We claim no rights over the your generated contents." Newer Wan versions (2.5 and later, including Wan 3.0) have no open weights on Wan-AI's Hugging Face page as of October 11, 2026; the newest uploads there are Wan 2.2 Animate 2 (August 2026, with defaults its card says are tuned for eight A800 GPUs) and Wan-Dancer, both Apache 2.0, and ComfyUI offers the later versions as paid API nodes. Our Wan 3.0 status page tracks that. For a character-animation workflow, see our Wan 2.2 Animate tutorial.
MiniMax H3 is the one to read twice. Its weights run natively in ComfyUI and generate 4- to 15-second clips with stereo audio, but Comfy's docs state that commercial use of locally generated outputs requires a MiniMax commercial license, while generations on Comfy Cloud include commercial rights. Our MiniMax H3 guide covers the download and the license.
How much VRAM do you need?
Vendor and Comfy statements, not our measurements. They mix "fits within" guidance with measured use, so read them as a rough order. Source: Model cards, BFL GitHub and docs.comfy.org, checked October 2026
- Around 8 GB: SDXL and its fine-tunes, Wan 2.2 5B with ComfyUI's offloading, and FLUX.2 [klein] 4B by BFL's lower estimate.
- Around 16 GB: Z-Image Turbo, and HiDream-I1 in fp8, which Comfy lists as needing more than 16 GB.
- Around 24 GB: Qwen-Image in fp8, the original FLUX.1 weights, and Wan 2.2 14B workflows in fp8.
- Data-center GPUs: full-precision FLUX.2 [dev] and Wan 2.2 A14B as run by the vendors' own scripts.
ComfyUI stretches these limits. Its README says it can run "even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights, and fp8, int8, nvfp4 or GGUF files trade some precision for smaller downloads and memory use. Expect slower runs the further you push below a model's comfortable size.
Getting these models into ComfyUI
The simplest route is a template: open Templates in the sidebar, pick a workflow for the model, and ComfyUI prompts you to download any missing files. Manual downloads go in ComfyUI/models: all-in-one checkpoints such as SDXL in checkpoints, standalone models such as Flux, Qwen-Image, Z-Image, Krea 2 and Wan in diffusion_models, plus their text encoders and VAEs. Our ComfyUI installation guide covers folders and requirements, and our ComfyUI-Manager guide covers the custom nodes some workflows need.
- 1Open Templates
The sidebar icon, or Workflow, then Browse Workflow Templates.
- 2Search the model name
For example Z-Image, Flux.2 Klein or Krea-2. If it is missing, update ComfyUI.
- 3Let it fetch the files
ComfyUI checks for missing models and prompts you to download them. Comfy Desktop downloads them for you.
- 4Accept gated licenses first
FLUX.2 [dev], FLUX.2 [klein] 9B, LTX-2.5 and Ideogram 4 are gated on Hugging Face: accept the terms there before the download works.
- 5Or place files by hand
Checkpoints in models/checkpoints; split models in models/diffusion_models, with their files in text_encoders and vae.
- 6Render once before changing anything
A clean image from the unmodified template proves the files, and your VRAM, are fine.
New to the node graph? Our ComfyUI learning curve guide explains what to learn first and what to ignore.
Picks for AI influencer workflows
- Consistent character stills: train a character LoRA on a base whose license fits your plans, such as Z-Image, FLUX.2 [klein] 4B, Qwen-Image or SDXL. Our LoRA training guide walks through dataset and settings; no LoRA holds the same face in every image, so review each output.
- Edits and outfit changes: Qwen-Image-Edit or FLUX.2 [klein] 4B for commercial projects; FLUX.1 Kontext [dev] for personal ones.
- Fast drafts: Z-Image Turbo; see our Z-Image Turbo tutorial.
- Motion: animate finished stills with Wan 2.2 image-to-video or Wan 2.2 Animate.
- Read the license on the exact file, not the family: fine-tunes and larger variants often differ from the base
- Check any revenue cap against your business: Krea 2 and SD 3.5 set theirs at $1M, LTX at $10M
- Match the file to your VRAM: fp8, int8 and nvfp4 files are smaller than bf16
- Decide which base your character LoRA will be trained on, because a LoRA only works on its own model family
- Prefer .safetensors files from the vendor or the Comfy-Org repositories
- Start from the official template so the sampler settings match the model
- Follow each platform's AI-disclosure rules, and only create likenesses you have the right to use
Build a consistent AI persona with the right model stack
Picking a model is one decision in a longer pipeline: character LoRA training, consistent images across many generations, animation and publishing. Our AI Influencers program covers that workflow: its Flux and SDXL Setup lesson walks through base-model choice and templates, and the Consistent Faces (LoRA Training) and Body Consistency (ControlNet) lessons build the character on top.
Best ComfyUI models FAQ
What is the best ComfyUI model in 2026?
It depends on the job and the license you need. For fast, commercially usable images, start with Z-Image Turbo or FLUX.2 [klein] 4B, both Apache 2.0. For text inside images and editing, use Qwen-Image and Qwen-Image-Edit. For anime, SDXL fine-tunes such as Illustrious-XL still have the largest LoRA ecosystem. For open video, Wan 2.2 is the default choice.
What is the best ComfyUI checkpoint for realistic people?
For a persona you will post commercially, Z-Image and FLUX.2 [klein] 4B are the safest bases: both are Apache 2.0, both fit consumer GPUs, and both have an undistilled variant for training a character LoRA. Krea 2 is a newer option whose license allows commercial use only under $1 million in annual revenue. SDXL fine-tunes such as Juggernaut XL v9 and RealVisXL V5.0 still suit 8 GB cards. We have not scored their output side by side.
Which ComfyUI models can I use commercially?
Apache 2.0 models allow it: Z-Image, FLUX.2 [klein] 4B, FLUX.1 [schnell], Qwen-Image and its Edit versions, Chroma1-HD and Wan 2.2. HiDream-I1 is MIT and SDXL uses CreativeML Open RAIL++-M. SD 3.5, Krea 2, LTX-2 and Pony V7 allow it below revenue or use limits. FLUX.1 [dev], FLUX.1 Krea and Kontext [dev], FLUX.2 [dev], FLUX.2 [klein] 9B, Qwen-Image-2.1, Ideogram 4 and NoobAI-XL are non-commercial without a separate license.
Is FLUX.1 [dev] free for commercial use?
Not as a model. Its license allows use only for non-commercial purposes and says revenue-generating use is not one. The Outputs clause lets you use outputs "for any purpose (including for commercial purposes), except as expressly prohibited herein", but the Restrictions section also bars using data produced by the model for commercial or production purposes. For paid work, get a license from Black Forest Labs or use an Apache 2.0 model.
How much VRAM do I need for Flux in ComfyUI?
The original FLUX.1 [dev] file is about 23 GB, and Comfy suggests Krea's original weights only if you have about 24 GB of VRAM. FLUX.2 [klein] 4B needs far less: Black Forest Labs quotes about 8 GB on GitHub and 13 GB on the model card, and Comfy measured 8.4 GB for the distilled version. Full FLUX.2 [dev] needs an H100-class GPU unless you use a quantized build.
What is the best ComfyUI model for anime?
SDXL fine-tunes such as Illustrious-XL and NoobAI-XL, plus the AuraFlow-based Pony V7, because most anime LoRAs target them. Check each checkpoint's license: NoobAI-XL prohibits any commercialization, and community fine-tunes set their own terms, so read the model page for the exact file you download.
What is the best open video model for ComfyUI?
Wan 2.2, released under Apache 2.0. Its 5B model fits on about 8 GB of VRAM with ComfyUI's offloading, according to Comfy's docs, while its 14B models need much more memory. LTX-2.5 adds synchronized audio under a community license, and MiniMax H3 needs a commercial license for locally generated outputs. Wan 2.5 and later, including Wan 3.0, had no open weights as of October 11, 2026.
Related reads
Free · no card needed
Start with the free Starter Kit.
An AI character bible from AI Influencers, plus four more working files.
Model picked. Now build the persona on it.
The AI Influencers program inside All Access starts where this page ends: base-model setup in ComfyUI, LoRA training for a consistent face, ControlNet, photorealism and video. All four programs, live coaching and the private community come in the same subscription.
Get the next model update when it lands
This list changes whenever a new open model ships. The free Telegram channel carries the updates and the guides that follow them.