Models and the Model Hub
On this page
- Smart model detection and architecture presets
- Availability on the connected server
- Misfiled models still load
- Model information
- GGUF quantized models
- INT8-Fast quantized models
- FLUX 3 Image Klein
- MiniMax-H3-Image
- Qwen-Image 2.1 and Qwen-Image 2.1 Turbo
- Architecture registry
- Flux.2 Klein
- Flux.1 and SD3 from split files
- Stable Cascade
- Wan as an image model
- Kolors
- Ming-Image 0.1
- MOSS-Image
- PixArt Sigma and PixArt Alpha
- Krea 2 uncensored mode
- Recommended models (auto-download)
- Model Hub (CivitAI / HuggingFace browser)
- CivitAI-enriched galleries
- Bulk CivitAI scan
- YuE2 music checkpoints
- Model manager
- Model requests (multi-user)
#Smart model detection and architecture presets
MooshieUI recognizes a broad set of model architectures and auto-applies an optimal sampler, scheduler, steps, CFG, and resolution for each. Families include (among others): SD 1.5, SD 2.x, SDXL, Illustrious/NoobAI, Pony Diffusion, SD3/SD3.5, the Flux family (Flux.1 Dev/Schnell/Krea, Flux 2 Klein variants and FLUX 3 Image Klein), MOSS-Image, Chroma, Z-Image, Wan, Qwen, Krea 2, Ideogram 4 (Flux2 VAE path), AuraFlow, PixArt, HunyuanDiT, Stable Cascade, Kolors, Ming-Image, Anima (COSMOS), Mugen (SDXL + Flux2 VAE), and Nanosaur (1.2B DiT).
The instruction-editing families (Qwen Image Edit, Qwen Image Edit Plus, Flux.1 Kontext, Flux.2 Klein, FLUX 3 Klein, MiniMax-H3-Image, MOSS-Image and Ming-Image Design) are detected too, and unlock Image Edit Mode when loaded.
Behavior worth knowing:
- Accelerated models (turbo, lightning, lcm, hyper in the name) get reduced steps and lower CFG automatically.
- Rectified flow models inject the correct ModelSampling node with architecture-specific shift values.
- FluxGuidance is injected for Flux Dev; Flux Schnell is detected and skipped.
- Hash-based identification - models are recognized by SHA256 hash (CivitAI AutoV2), so renamed files still resolve.
- Transparent backgrounds - images with a real alpha channel (for example from NovelAI V5 Transparent BG) are saved with their transparency instead of being made opaque.
- Structural identification - MooshieUI inspects the safetensors tensor keys and identifies the architecture from the weights themselves, and when those weights clearly disagree with the filename the weights win (so an SDXL model misnamed like a Flux one is still classified as SDXL). Anima (COSMOS), Cosmos Predict2, Wan 2.1, Flux, Qwen-Image, SD3, SDXL, and SD 1.5 are all recognized this way, whatever the file is named, so an Anima fine-tune renamed without an
anima/yumetoken (for example PerfectDeliberate) or a quantized FP8 build is still classified as the Anima family. An Anima fine-tune that ships its own text encoder named after the model (for examplename_txt.safetensorsbesidename.safetensors) gets that encoder picked automatically. This matters twice over: it selects the matching text encoder for split models, and it routes ControlNet through the Anima LLLite path where that is required. See ControlNet and Style Transfer.
#Availability on the connected server
Generation pickers use filenames advertised by the connected ComfyUI server. Recommended entries distinguish Available in ComfyUI, Downloaded locally, unavailable to ComfyUI, and Unavailable in ComfyUI. A file on the desktop is not automatically available to a remote worker. Refresh after installing the required files on that server; selecting a missing remote model does not start a local download.
When the connection changes, previous server inventory and stale refresh results are discarded. Renamed or nested files resolve to an advertised server filename; ambiguous basenames remain separate choices.
#Misfiled models still load
ComfyUI's own loaders only accept files from their own folder, so a diffusion model (unet or DiT only) dropped into the checkpoints folder, or a full checkpoint sitting in diffusion models, used to fail at generation time. MooshieUI now reads the weights to tell which kind a file really is, switches to the matching loader without asking, and loads the file by its full path. Misfiled models simply work, with nothing to move and no error to decipher.
Quantized GGUF files are the exception, since their loader comes from a third-party node that takes no path. Keep those in the folder their category expects.
A component in the wrong picker is different, because it cannot be loaded as something else. Text encoders and VAEs on their own cannot generate images, so when one is picked as the model, MooshieUI shows a red error under the picker that names the file, says what it is, and says where it goes (for example under Text Encoder or VAE). The same check covers a text encoder or model picked under VAE, and a VAE or model picked under Text Encoder, and Generate stops with the same message instead of failing inside ComfyUI. Detection reads the file's tensor layout, GGUF text encoders included, so renaming a file does not change the result.
#Model information
Selecting a model shows what the file itself declares: the preview thumbnail stored inside it, plus Date, Implementation, SHA256, Merged from, Preprocessor, Encoder layer and ModelSpec version where the author filled them in. An architecture worked out from the weights rather than declared in the file is marked inferred. Anything else the file declares is listed under Other ModelSpec fields, so nothing a model author wrote is hidden.
#GGUF quantized models
.gguf diffusion models and text encoders generate like any other model, so you can run large architectures (Krea 2 and similar) on cards that cannot hold the full-precision weights.
- Turn on Use Split Model Components, then pick the
.gguffile under Diffusion Model. Quantized encoders work the same way under Text Encoder, which lists files from both thetext_encoders/andclip/folders. - Routing is automatic. ComfyUI's core loaders cannot read
.gguf, so MooshieUI sends those files to the ComfyUI-GGUF loaders instead. There is nothing to configure. - The ComfyUI-GGUF nodes install during ComfyUI setup and are re-checked on every start. If the install fails, ComfyUI still starts normally and only GGUF models are unavailable until it succeeds.
- Detection works on GGUF files too. The family now resolves from the GGUF header itself (its architecture field and tensor names) as well as the filename, sidecar metadata, and hash, so a renamed GGUF still lands on the right family; the accelerated variant and recommended encoder round out the preset. A file like
Krea-2-Turbo-Q5_K_S.ggufgets the correct settings instead of falling back to unknown. Wulver (a Krea 2 fine-tune whose filenames carry no Krea marker) is also detected: files withoutnon_turboin the name are classified as Krea 2 Turbo, andnon_turbofiles get the standard Krea 2 defaults.
Put .gguf diffusion models in your diffusion models folder and quantized encoders in text_encoders/. The model manager (below) handles both categories.
#INT8-Fast quantized models
INT8 W8A8 and ConvRot builds (Flux 2, Flux 2 Klein, Z-Image, Chroma, Krea 2, Qwen-Image, Anima and Ideogram 4) need the ComfyUI-INT8-Fast loader rather than ComfyUI's core loader. It is NVIDIA-only.
- Turn on Use Split Model Components, pick the INT8 file under Diffusion Model, then turn on the INT8-Fast loader toggle. Enable ConvRot underneath is on by default and should stay on for ConvRot files.
- The node pack is not installed by default. The toggle shows Install ComfyUI-INT8-Fast the first time; installing restarts ComfyUI.
- If a filename looks like an INT8 or ConvRot build while the toggle is off, MooshieUI points this out under the model picker. A generation that fails with
unknown quantization format for layeris the same problem. - The toggle is disabled (greyed out) when the selected diffusion model is a
.gguffile. The tooltip reads "Not available for .gguf models (already optimized)" to explain why. Use the GGUF loaders for those files instead.
#FLUX 3 Image Klein
FLUX 3 Image Klein comes as a distilled model and a Base model. Both are detected from BFL and community file names, CivitAI base model names and the file's own tensors, so a renamed file still gets the right preset: 16 steps at CFG 1 for distilled, 50 steps at CFG 3.5 for Base, Euler on simple. Klein needs the Qwen3-VL 4B text encoder (the same one Krea 2 uses) and the Flux 2 VAE. If either is missing it downloads on its own, and generating with a different encoder or VAE stops with a message naming the right one. Klein also works in Image Edit Mode with up to 10 references. It needs ComfyUI v0.41.0 or newer.
The Model Hub quick links include FLUX 3 Klein INT8 (HF access) and FLUX 3 Klein Base INT8 (HF access). These come from BFL's gated Hugging Face repo, so they need a Hugging Face token from an account that has accepted the FLUX license. ComfyUI loads these INT8 ConvRot files itself, so leave the INT8-Fast loader off for FLUX 3.
#MiniMax-H3-Image
MiniMax-H3-Image is the 33B MiniMax H3 image model. Txt2img and Image Edit (up to 16 references) share one checkpoint. It uses the same Qwen3-VL-32B text encoder as the H3 video model, so a copy already downloaded for video is reused automatically.
- Default preset: 6 steps at CFG 1, using the model's own sigma grid.
- pDMD 2-step mode: a separate LoRA (
minimax_h3_pdmd_2nfe_comfyui.safetensors) or a checkpoint with it merged in reduces generation to 2 steps. The model is detected as the pDMD family and the preset is applied automatically. - Resolution: works best from 4 to 20 MP (native around 16 MP, 4096x4096). Batch size is always 1 (the model's own limit); larger batches run as sequential samplers.
- Model Hub: INT8 and W4A8 downloads are available.
- License: the MiniMax-H3-Image Community License. Written permission from MiniMax is needed above US$1M annual revenue or to offer it as a hosted service. A notice appears under the model picker when this model is selected.
#Qwen-Image 2.1 and Qwen-Image 2.1 Turbo
Qwen-Image 2.1 and Qwen-Image 2.1 Turbo are detected as a distinct architecture from Qwen-Image 1 (earlier releases misidentified them and they could not generate). Both use the right encoder and VAE automatically.
- Qwen-Image 2.1 is detected from its tensor layout and the 2.1 marker in the filename.
- Qwen-Image 2.1 Turbo runs on its own 8-step schedule, which the preset applies automatically when the Turbo variant is detected.
- Both are available in Image Edit Mode with up to 10 references.
- Downloads are available from the Model Hub quick links.
#Architecture registry
From v2.4.0, model families are declared once in a registry that sets detection, the text encoder and VAE, extra model files, supported modes, sampling defaults and Model Hub downloads. A family on the registry downloads its missing encoder and VAE on its own and refuses a mode it cannot run with a message, instead of producing a broken image. The families below moved onto it in v2.4.0.
#Flux.2 Klein
Flux.2 Klein 4B and 9B, distilled and Base, run txt2img and Image Edit Mode with up to 10 reference images. Defaults follow ComfyUI's bundled Klein templates: 4 steps at CFG 1 for distilled, 20 steps at CFG 5 for Base. A Klein 9B KV file (klein, 9b and kv in the name) also gets the KV cache. Klein is told apart from Flux.1 by its tensors, so a renamed file is no longer read as Flux.1. The Model Hub links BFL's FP8 files; the 9B repos are gated and need a Hugging Face token.
#Flux.1 and SD3 from split files
Split Flux.1 files (dev, schnell, Krea, Kontext) load T5-XXL and CLIP-L together, as ComfyUI does. Earlier releases loaded T5 alone, which degraded every image without an error. CLIP-L and the Flux VAE download when missing.
SD3 and SD3.5 split files load CLIP-L, CLIP-G and T5-XXL together, and Stability's own release files (sd3_medium, sd3.5_large, sd3.5_medium, sd3.5_large_turbo) work as they are. SD3.5 Large Turbo runs at 4 steps, CFG 1. The encoders download from Comfy-Org's ungated repo; the SD3 VAE is only in Stability's gated repos, so a GGUF model without one stops with a message pointing at its Model Hub entry.
#Stable Cascade
Stable Cascade generates with the official two-stage workflow: stage C at your settings, then stage B (10 steps, CFG 1.1) and stage A to decode. Txt2img, img2img, hires fix and face fix are supported. Stage B is the official all-in-one ComfyUI checkpoint, or a stage B UNet plus stage A; the model panel uses one you have or downloads the checkpoint (about 4.6 GB). ControlNet and pause/resume are not available for Cascade.
#Wan as an image model
Wan 2.1 (1.3B and 14B), Wan 2.2 A14B and Wan 2.2 TI2V 5B generate stills with the umT5 text encoder. Each gets its own VAE (TI2V 5B uses the 48-channel Wan 2.2 VAE), and Wan 2.2 A14B runs its high-noise and low-noise experts in turn.
#Kolors
Kolors runs through a loader node bundled with MooshieUI, so no third-party node pack is needed. It needs the Kolors UNet, the Kolors VAE and all shards of the ChatGLM3 text encoder; the Model Hub entry downloads the full set (about 17.8 GB). The tokenizer downloads once on first use. ControlNet is not available for Kolors.
#Ming-Image 0.1
inclusionAI's Ming-Image comes as two models:
- Ming-Image Design: txt2img, img2img, inpainting and Image Edit Mode with up to 8 references. Defaults are 12 steps at CFG 1 and 2048x2048, so the negative prompt box is hidden. A Transparent background checkbox under Dimensions asks for a cut-out, and the PNG keeps its transparency. Upscaling, face fix and segment detail drop the alpha channel.
- Ming-Image Layer: a separate checkpoint with its own text encoder.
#MOSS-Image
OpenMOSS's MOSS-Image comes as MOSS-Image (SFT) and MOSS-Image Base, for txt2img, img2img and Image Edit Mode. Both are detected from Comfy-Org and community file names, CivitAI base model names and the file's own tensors; only a base marker in the name selects Base, since both share the same tensors. The preset follows Comfy-Org's templates: 25 steps at CFG 2.5, er_sde on simple, 1024x1024, with the negative prompt kept on. MOSS-Image needs the Qwen3.5 9B text encoder and the Flux 2 VAE. If either is missing it downloads on its own, and generating with a different encoder or VAE stops with a message naming the right one. It needs ComfyUI v0.43.0 or newer, which the managed ComfyUI uses from v2.3.34.
The Model Hub quick links include MOSS-Image INT8, MOSS-Image Base INT8 and Qwen3.5 9B W4A8 (MOSS-Image), from Comfy-Org's ungated repo. ComfyUI loads these INT8 files itself, so leave the INT8-Fast loader off for MOSS-Image.
#PixArt Sigma and PixArt Alpha
Both models now generate (earlier releases detected them but could not generate). They always load split: the DiT through the diffusion model loader and the T5-XXL text encoder through the text encoder picker. No custom nodes are needed; ComfyUI's core loaders handle both the original .pth keys and diffusers safetensors format.
- Text encoder: T5 v1.1 XXL (
t5xxl_fp16ort5xxl_fp8_e4m3fn). MooshieUI downloads the FP8 version automatically when it is missing. - VAE: PixArt Sigma uses the SDXL VAE (
sdxl_vae.safetensors); PixArt Alpha uses the SD 1.5 VAE. Both download automatically when missing. - Preset: 20 steps, CFG 4.5, DPM++ 2M sampler, normal scheduler, 1024x1024.
- Detection: Sigma and Alpha are told apart by their tensor structure (Sigma has no micro-conditioning weights; Alpha does), so renaming a file does not confuse them.
- Prompt style: natural language (full sentences, not tags).
#Krea 2 uncensored mode
Krea 2 reads your prompt through Qwen3-VL-4B, a chat model trained to refuse some requests, and its TextFusion layers (which turn that encoder output into conditioning for the image model) learned to tone those concepts down. Things can come out weakened or missing, and not only NSFW ones. When a Krea 2 model is selected, an Uncensored mode panel appears under the Text Encoder picker.
- Uncensored mode combines two fixes for the two stages: an abliterated (Heretic) copy of the same encoder at the same fp8 precision (DreamFast/Qwen3-VL-4b-Heretic-ComfyUI), and Capitan01R's TextFusion refusal-reduction LoRA (v2.0 full rank) at strength 1.0.
- Download and turn on fetches whatever is missing (about 7.4 GB for both) and switches the mode on when the downloads finish. If you set up the uncensored encoder in v2.3.16, the panel shows Finish setup to add the LoRA (2.6 GB). You also get a one-time notification with a Download and turn on button that does the same; download progress shows in the download banner.
- Turn on and Turn off switch between uncensored mode and the standard encoder. The panel shows Active while it is on. The choice is remembered across models and restarts, and picking an encoder by hand in the Text Encoder picker updates it too.
- The LoRA is applied automatically to Krea 2 generations only and does not appear in your LoRA list, so it never carries over to another model family. It is skipped if you already have a version of it enabled yourself, and it is recorded in the image metadata.
- It cannot add concepts the image model itself never learned. If something still does not appear in uncensored mode, the limit is in Krea 2's own training.
#Recommended models (auto-download)
In managed local mode, a few curated checkpoints can auto-download on selection with a live progress bar and file size, including a NoobAI/Illustrious checkpoint Anima Base v1.0 (~5.6 GB), and Anima Turbo and Aesthetic v1.1. The exact recommended list evolves between releases, so treat the in-app list as the source of truth.
#Model Hub (CivitAI / HuggingFace browser)
Browse and download models from inside the app:
- Filter by type (Checkpoint, LoRA, ControlNet, Upscaler, VAE, Textual Inversion), sort (Highest Rated, Most Downloaded, Newest), time period, architecture, and file format (SafeTensor, PickleTensor, GGUF, Diffusers, CoreML, ONNX, Other).
- Add a CivitAI API key for gated content.
- Open on CivitAI opens the model's page. Mature models open on civitai.red, where CivitAI serves them; pasted civitai.red model links work too.
- One-click download routes the file to the correct folder with a real-time progress bar.
Switch the source from CivitAI to Direct URL to install from any legal direct model link (Hugging Face, a mirror, a private host). Paste the Model URL, confirm the Filename, and choose where it lands under Category: Checkpoint, Diffusion Model, LoRA, Upscaler, VAE, ControlNet, Textual Inversion, or Text Encoder. Diffusion Model is the right home for a bare DiT checkpoint that carries no built-in text encoder. Hugging Face links must be /resolve/main/ file URLs, not model page URLs; the field tells you if you paste the wrong one. Quick links cover common downloads, including SDXL Base 1.0, the OmniSR upscalers, the Qwen3-VL 4B FP8 encoder that Krea 2 uses, and Krea 2 - Wulver v0.5 FP8 (Turbo).
#CivitAI-enriched galleries
The bottom-panel LoRA and Checkpoint galleries auto-fetch CivitAI metadata (name, preview images, model ID) for your local files and cache it for 24 hours. Their CivitAI links open mature models on civitai.red. The LoRA gallery lets you set a thumbnail from a CivitAI image URL. Both support search and sort by name, date, or family, plus right-click actions. The sort dropdown also has Folder and Tree modes that group your LoRAs by their on-disk folders (Tree mirrors the Model Manager's collapsible folder view), handy when your LoRAs are organised into subfolders.
#Bulk CivitAI scan
Above the search bar in the LoRA gallery, a Scan all models on CivitAI button hashes every LoRA, checkpoint, diffusion model, and embedding in your library, queries CivitAI by hash, and writes a .civitai.info sidecar file next to each match. This populates the gallery thumbnails, names, and model IDs in one pass rather than waiting for them to be fetched on demand.
While the scan runs, a progress bar shows the current file being hashed ("Hashing filename...") and a "{current} / {total}" counter. A Cancel button stops the scan at any point (it drains to "Cancelling..." before halting). When complete, a summary line reports how many files were found on CivitAI, how many were not found, and how many were skipped.
Files that already have a cached .civitai.info sidecar are skipped by default. Once a scan has finished, the button label changes to Re-scan all models, which forces a fresh hash and lookup even for already-cached files. Use Re-scan all models if a model's CivitAI page was updated after your original scan. A Dismiss button closes the progress panel after the scan completes.
#YuE2 music checkpoints
Music has a separate YuE2 checkpoint selector under Generation settings. Models and setup offers the official INT8 (4.0 GB) and BF16 (7.8 GB) downloads, subject to the account's download permissions. These files go in ComfyUI/models/checkpoints and include the generator, text encoder and audio VAE. The weights have a non-commercial license. See Music Generation for runtime and hardware requirements.
Native covers also need SheetSage2 under the separate Audio encoder selector. Its download is 1.39 GB and belongs in the transcription worker's models/audio_encoders folder. It is not a YuE2 checkpoint. See Music Generation for transcription and score review.
#Model manager
In Settings you can manage local model files across categories (checkpoints, loras, vae, upscale, controlnet, embeddings, ultralytics, text encoders, diffusion models): search by filename, delete files, move files between configured directories or to a different category (for example a checkpoint into your diffusion models folder), and see total disk usage per category.
Switch between List and Tree view: List is a flat, sorted list, while Tree groups files by the folders they're stored in so large collections mirror your actual on-disk folder structure. You can create folders (New Folder) and move models into them.
#Model requests (multi-user)
In hosted/multi-user deployments, non-admin users can request a model download by CivitAI URL. Admins and moderators see a pending/approved/denied queue and can approve (which starts the download) or deny with a reason. See Server, LAN and Multi-User.