Nothing is downloaded at startup. A checkpoint is fetched by the first run that needs it, and
only with features.network: true. With that off, the run stops and names the file and every
folder it searched.
Published at https://huggingface.co/WAS/EMA-VFI. Goes in ComfyUI/models/EMA-VFI.
| File | Size | Quality | Multiplier above 2 |
|---|---|---|---|
ours_t.safetensors |
263 MB | best | yes |
ours.safetensors |
263 MB | best | no |
ours_small_t.safetensors |
58 MB | lower | yes |
ours_small.safetensors |
58 MB | lower | no |
A _t file is the one to take: it lands a frame anywhere between two, so a multiplier above 2
works.
Video Motion Model Loader builds the network that Video Motion and Video Temporal
Upscale read. Files go in ComfyUI/models/optical_flow, or a folder an extra_model_paths.yaml
entry names under the key optical_flow. With features.network: true, a file in this table is
fetched by the first run that picks it.
| Save as | Network | Size | Download |
|---|---|---|---|
sea_raft_m_spring.safetensors |
SEA-RAFT M, the default | 79 MB | model.safetensors from MemorySlices/Tartan-C-T-TSKH-spring540x960-M |
sea_raft_s_spring.safetensors |
SEA-RAFT S, about 1.3x faster | 36 MB | model.safetensors from MemorySlices/Tartan-C-T-TSKH-spring540x960-S |
sea_raft_m_ct.safetensors |
SEA-RAFT M, general purpose | 79 MB | model.safetensors from MemorySlices/Tartan-C-T432x960-M |
flowseek_t_ct.safetensors |
FlowSeek T, reads depth too, about 1.5x slower | 171 MB | optical_flow/flowseek_t_ct.safetensors from WAS/was-node-suite-weights |
flowseek_t_tskh.safetensors |
FlowSeek T, the TSKH training | 171 MB | optical_flow/flowseek_t_tskh.safetensors from WAS/was-node-suite-weights |
Every SEA-RAFT repository names its file model.safetensors. A copy downloaded by hand is saved
under the name in the first column. Upstream's own .pth releases load under the names they are
published with: SEA-RAFT's Tartan-*.pth from princeton-vl/SEA-RAFT
and FlowSeek's flowseek_*.pth from mattpoggi/flowseek.
| Weights | Licence |
|---|---|
| SEA-RAFT | BSD-3-Clause, Princeton Vision & Learning Lab |
| FlowSeek, with the Depth Anything V2 weights inside it | Apache-2.0 |
| File | Goes in | Size | Read by | Download |
|---|---|---|---|---|
minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors |
ComfyUI/models/latent_upscale_models |
691 MB | H3 Upscale Video, upscale_model |
minimax_h3_latent_upscaler_3d_conv_v1/ in LBH-123-AI/Minimax_h3_latent_Upscaler |
taeh3.safetensors |
ComfyUI/models/vae_approx |
22,709,752 bytes | H3 Extend Window, live_preview, drawn in the Prompt Timeline |
safetensors/taeh3.safetensors in madebyollin/taehv |
The upscaler is fetched by the first run that picks it with features.network: true, and the
_bf16 file beside it loads too. taeh3.safetensors is never fetched. Another file named
taeh3, the older per-frame decoder, is a different size and does not load. Both folders are
read through extra_model_paths.yaml only where a root carries an entry under the key
latent_upscale_models or vae_approx.
| Weights | Licence |
|---|---|
| The H3 latent upscaler | Apache-2.0, LBH-123-AI |
taeh3.safetensors |
MIT, Ollin Boer Bohan |
hdrcnn.safetensors,
118 MB, from hdr/ in WAS/was-node-suite-weights,
into ComfyUI/models/hdr. models/hdr/hdr/ is read too, and an extra_model_paths.yaml entry
under the key hdr.
Searched in order before anything is fetched: ComfyUI/models/<family>, Hugging Face's own
cache, then the ckpts directory of any pack under custom_nodes. Nothing is copied or moved.
A model is held for the life of the process, so one loader feeds as many nodes as wanted at the
cost of one set of weights.
| Family | Folder |
|---|---|
| Depth | ComfyUI/models/depth_anything |
| Pose | ComfyUI/models/pose |
| Segmentation | ComfyUI/models/segmentation |
| Edge | ComfyUI/models/hed, pidi, teed |
| Lineart | ComfyUI/models/lineart, lineart_anime, manga_line |
| Line segments | ComfyUI/models/mlsd |
| AnyLine | ComfyUI/models/teed |
| Intrinsics | ComfyUI/models/intrinsics |
| Denoising | ComfyUI/models/denoise |
| Low light | ComfyUI/models/low_light |
| Model | Answers | Size |
|---|---|---|
Depth Anything V2 Small |
Depth | 99 MB |
Depth Anything V2 Base |
Depth | 390 MB |
Depth Anything V2 Large |
Depth | 1.3 GB |
DPT SwinV2 Tiny |
Depth, the quickest here | 164 MB |
DPT Large |
Depth, a heavier reading than Depth Anything | 1.4 GB |
ViTPose Small |
Body pose | 130 MB |
ViTPose Wholebody |
Body pose with feet, face and both hands, 133 points | 97 MB |
ViTPose Animal |
Animal pose, 17 points | 97 MB |
ViTPose Base |
Body pose | 380 MB |
SegFormer B0 ADE20K |
150 classes per pixel | 15 MB |
SegFormer B2 ADE20K |
150 classes per pixel | 110 MB |
SegFormer B4 ADE20K |
150 classes per pixel | 260 MB |
HED Soft Edge |
Soft edges | 29 MB |
PiDiNet Soft Edge |
Soft edges, crisper than HED | 3 MB |
TEED Soft Edge |
Soft edges, the lightest of the three | 2 MB |
Lineart |
Drawn lines, fine | 62 MB |
Lineart Coarse |
Drawn lines, heavier | 62 MB |
Lineart Anime |
Drawn lines, trained on anime | 218 MB |
Manga Line |
Manga and comic linework | 173 MB |
MLSD Line Segments |
Straight runs, for architecture and interiors | 6 MB |
AnyLine |
A fine edge pass merged with a lineart pass | 0.2 MB |
Marigold IID Appearance |
Albedo, roughness, metallicity and material | 1.9 GB |
Marigold IID Lighting |
Albedo, shading and residual | 1.9 GB, or 1.7 GB once the other is on disk |
SCUNet |
Denoising, the widest range of noise | 72 MB |
NAFNet SIDD width32 |
Denoising, light sensor grain | 117 MB |
NAFNet SIDD width64 |
Denoising, the same at greater width | 464 MB |
DarkIR |
Low light, and takes out blur and noise with it | 13 MB |
Retinexformer NTIRE |
Low light, the general choice | 6.5 MB |
Retinexformer LOL v1 |
Low light | 6.5 MB |
Retinexformer LOL v2 Real |
Low light | 6.5 MB |
Retinexformer LOL v2 Synthetic |
Low light | 6.5 MB |
Retinexformer FiveK |
Low light, the gentlest lift | 6.5 MB |
Retinexformer Extreme Dark |
Low light, the hardest lift | 6.5 MB |
Retinexformer Dark Motion |
Low light, moving subjects | 6.5 MB |
Retinexformer Indoor Night |
Low light, interiors | 6.5 MB |
Retinexformer Outdoor Night |
Low light, exteriors | 6.5 MB |
HVI-CIDNet Generalization |
Low light, a second opinion | 8 MB |
HVI-CIDNet FiveK |
Low light, photographic tone | 8 MB |
HVI-CIDNet SICE |
Low light, mixed exposure | 8 MB |
HVI-CIDNet Extreme Dark |
Low light, near black | 8 MB |
The six intrinsic maps, albedo, roughness, metallicity, material, shading and residual,
come from Marigold. They run in torch on the published safetensors and need nothing
installed. Both repositories publish the same autoencoder, and it is fetched once: the
second checkpoint reads the copy the first one left in ComfyUI/models/intrinsics.
Marigold v2 is a model for depth_map, normal_map and albedo, and the one model the
node does not download. Picking it reads the model, vae and conditioning_name
inputs instead. Take the files from
Comfy-Org/marigold-v2-0 and place them
yourself:
| File | Goes in | Size |
|---|---|---|
qwen_image_edit_2509_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
20.5 GB |
marigold_v2_depth_log_stage2.safetensors |
ComfyUI/models/loras |
1.7 GB |
marigold_v2_normals.safetensors |
ComfyUI/models/loras |
1.7 GB |
marigold_v2_albedo.safetensors |
ComfyUI/models/loras |
1.7 GB |
marigold_v2_depth_log_stage2_vae.safetensors |
ComfyUI/models/vae |
254 MB |
marigold_v2_normals_vae.safetensors |
ComfyUI/models/vae |
254 MB |
marigold_v2_albedo_vae.safetensors |
ComfyUI/models/vae |
254 MB |
marigold_v2_depth_conditioning.safetensors |
ComfyUI/models/embeddings |
2.4 MB |
marigold_v2_normals_conditioning.safetensors |
ComfyUI/models/embeddings |
2.5 MB |
marigold_v2_albedo_conditioning.safetensors |
ComfyUI/models/embeddings |
2.6 MB |
Which map each answer takes:
| Answer | LoRA | VAE | Conditioning |
|---|---|---|---|
depth_map |
..._depth_log_stage2 |
..._depth_log_stage2_vae |
..._depth_conditioning |
normal_map |
..._normals |
..._normals_vae |
..._normals_conditioning |
albedo |
..._albedo |
..._albedo_vae |
..._albedo_conditioning |
Choosing Marigold v2 on the canvas wires Load Diffusion Model on the transformer into
model. A loader already in the graph on a qwen_image_edit_2509 file is reused; otherwise
one is placed beside the node, on qwen_image_edit_2509_int8_convrot.safetensors or, where
the folder lacks it, another qwen_image_edit_2509 build it holds. That is the only wire the answer needs: adapter_name,
vae_name and conditioning_name stay on auto and find that row's files by name, so one
transformer feeds every map and no LoRA or VAE loader is placed.
A file the folders lack is named on its widget instead of auto, and ComfyUI lists it under
missing models with its download. Choosing another model takes the loader's link off the
node and puts a named, missing file back to auto; changing the answer moves each widget to
that map's file.
The three overrides, where a file is named something else or a loader is wanted:
| Override | Use |
|---|---|
adapter_name set to already on the model |
A LoRA put on the transformer beforehand, at a strength of your own |
The vae socket |
A decoder from Load VAE or any node answering a VAE |
The conditioning socket |
A prompt embedding from a text encoder or a loaded conditioning |
conditioning_name stays on auto, which finds that map's embedding by name in
the embeddings folder. Naming a file overrides it. It is the text encoder's saved output, so
no text encoder is loaded.
The LoRA is what decides the map. Running the depth adapter against another map's embedding moves the result by 0.005 on a 0 to 1 scale, and against a short one by 0.226, so an override is worth setting only to a full length embedding.
One transformer serves all three answers, and only the LoRA, the VAE and the conditioning change. It occupies about 19 GB of VRAM, so a 24 GB card runs it with little room for a checkpoint beside it. The first answer costs about 10 seconds while the transformer loads, and each one after that about 0.8 seconds a frame.
| Trouble | Cause |
|---|---|
model is not wired |
No transformer reaches model, as in a workflow saved before it was wired. Set model_name to another model and back to Marigold v2, or wire Load Diffusion Model by hand. |
Value not in list on unet_name, adapter_name, vae_name or conditioning_name |
The file named there is not downloaded yet. Download it from the missing models list, or from the table above, and reload the page. |
found no adapter |
That map's LoRA is not in the loras folder, or the model already carries one and adapter_name was not set to already on the model. |
found no decoder |
That map's VAE is not in the vae folder. |
found no prompt embedding |
That map's conditioning file is not in the embeddings folder. |
| A map that looks like noise | The LoRA and the VAE name different maps. |
| Depth flat or inverted | The answer already turns log depth over, so no invert node is wanted after it. |
The conditioning menu is empty |
The files are not in ComfyUI/models/embeddings. Put them there and reload the page. |
Video Super Resolution (PS-SR) reads a checkout placed by hand. Nothing here is fetched,
whatever features.network is set to, and the group is off until pssr: true is set under
features: in config.yaml.
The checkout goes in ComfyUI/models/PS-SR. ComfyUI/models/ps-sr and ComfyUI/models/pssr
are read too, an extra_model_paths.yaml entry under any of those three names keeps it on
another drive, and PSSR_ROOT names a directory ahead of all of them.
| File under the checkout | Holds |
|---|---|
checkpoints/pretrained_models/base.safetensors |
The steady restoration pass |
checkpoints/pretrained_models/draft.safetensors |
The sharp restoration pass |
dependent_models/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth |
The Wan VAE |
models/Wan-AI/Wan2.1-T2V-1.3B/google/umt5-xxl |
The umt5 tokenizer |
Wan_SR/pipelines/ |
The two pipelines the node builds |
A checkout missing any of the first four is named in the error along with what is absent.
The transformer is not in that list. It comes in on the model input from Load Diffusion
Model, as Wan 2.1 T2V-1.3B or a finetune of it. The tagger, its LoRA and the 11 GB text
encoder are not read at all.
diffsynth is the one package the pack imports for this that ComfyUI does not already have.
The checkout's own requirements file names the rest.
ComfyUI's own loaders read every file, and nothing is fetched. The pack publishes them at
https://huggingface.co/WAS/was-node-suite-weights under models/kandinsky6/, laid out as
ComfyUI's models folder. The transformer files Kandinsky Lab publishes,
transformer/diffusion_pytorch_model.safetensors in each Kandinsky-6.0 repository, load as they
are.
| File | Goes in | Size | Loaded by |
|---|---|---|---|
kandinsky6_lite_distill_5s_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
4.2 GB | Load Diffusion Model |
kandinsky6_lite_distill_5s_w6a8.safetensors |
ComfyUI/models/diffusion_models |
3.8 GB | Load Diffusion Model |
kandinsky6_lite_5s_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
4.2 GB | Load Diffusion Model |
kandinsky6_lite_5s_w6a8.safetensors |
ComfyUI/models/diffusion_models |
3.8 GB | Load Diffusion Model |
kandinsky6_pro_distill_5s_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
34.9 GB | Load Diffusion Model |
kandinsky6_pro_distill_5s_w6a8.safetensors |
ComfyUI/models/diffusion_models |
29.4 GB | Load Diffusion Model |
kandinsky6_pro_5s_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
34.9 GB | Load Diffusion Model |
kandinsky6_pro_5s_w6a8.safetensors |
ComfyUI/models/diffusion_models |
29.3 GB | Load Diffusion Model |
qwen_2.5_vl_7b_fp8_scaled.safetensors |
ComfyUI/models/text_encoders |
9.4 GB | DualCLIPLoader, type kandinsky5 |
clip_l.safetensors |
ComfyUI/models/text_encoders |
246 MB | DualCLIPLoader, type kandinsky5 |
kandinsky6_audio_vae.safetensors |
ComfyUI/models/vae |
1.7 GB | Kandinsky 6 Audio VAE Loader |
hunyuan_video_vae_bf16.safetensors |
ComfyUI/models/vae |
493 MB | Load VAE |
hunyuan_video_vae_bf16.safetensors is published by Comfy-Org at
https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged under split_files/vae/.
| Transformer | Weights |
|---|---|
_int8_convrot |
8-bit, run as 8-bit |
_w6a8 |
6-bit, run with 8-bit activations, the smallest |
| Kandinsky Lab's own file | bfloat16 |
| Transformer | KSampler |
|---|---|
_distill_ |
10 steps, CFG 1, euler, simple |
| the others | 50 steps, CFG 5, euler, simple |
| File | Licence |
|---|---|
| The Kandinsky 6 transformers | MIT, Kandinsky Lab |
qwen_2.5_vl_7b_fp8_scaled.safetensors |
Apache-2.0, Alibaba Cloud |
clip_l.safetensors |
MIT, OpenAI |
kandinsky6_audio_vae.safetensors |
CC BY-NC 4.0, non-commercial use only, the MMAudio autoencoder in it; MIT, the BigVGAN vocoder |
hunyuan_video_vae_bf16.safetensors |
Tencent Hunyuan Community License, which does not apply in the European Union, the United Kingdom or South Korea |
| Group | Goes in | Widget | Repository |
|---|---|---|---|
sam |
models/sams |
ViT-H |
facebook/sam-vit-huge |
ViT-L |
facebook/sam-vit-large | ||
ViT-B |
facebook/sam-vit-base | ||
blip |
models/blip |
caption | Salesforce/blip-image-captioning-base |
| question | Salesforce/blip-vqa-base | ||
clipseg |
models/clipseg |
default | CIDAS/clipseg-rd64-refined |
midas |
models/midas |
DPT_Large |
Intel/dpt-large |
DPT_Hybrid |
Intel/dpt-hybrid-midas | ||
DPT_Small |
Intel/dpt-swinv2-tiny-256 | ||
diffusers |
models/diffusers |
any | any diffusers-format repository, kept as <owner>/<name> |
birefnet |
models/birefnet |
any | any full size BiRefNet .safetensors, published as WAS/was-node-suite-weights |
ben2 |
models/ben2 |
any | any BEN2 base .safetensors, published as ben2-base.safetensors |
Each is searched under its folder in these layouts, and Hugging Face's own cache after that.
| Layout | Example |
|---|---|
| The name alone | models/blip/blip-image-captioning-base/ |
| Owner and name | models/blip/Salesforce/blip-image-captioning-base/ |
| A Hugging Face cache tree | models/blip/models--Salesforce--blip-image-captioning-base/snapshots/<rev>/ |
The graphs in docs/workflows load these files through ComfyUI's own loaders. The Kandinsky 6,
Marigold v2, EMA-VFI and optical flow files they read are in those sections above. A graph names
some LoRAs inside a subfolder, as H3/taomate_h3_3step_comfy.safetensors: a file kept in that
subfolder loads as the graph names it, and one kept elsewhere is picked again from the loader.
Graphs: every minimax-h3-*.json and affine-minimax-h3-*.json.
| File | Goes in | Size | Download | Licence |
|---|---|---|---|---|
minimax_h3_fl2va_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
34.0 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_fl2va_pruned_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
21.0 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_fl2va_pruned_w6a8.safetensors |
ComfyUI/models/diffusion_models |
16.0 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_ref2va_pruned_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
21.0 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_ref2va_pruned_w6a8.safetensors |
ComfyUI/models/diffusion_models |
16.0 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_ref2va_silverturbo_pruned_int8_convrot.safetensors |
ComfyUI/models/diffusion_models |
21.0 GB | silveroxides/MiniMax-H3_tests | None stated |
singularity_minimax_h3_ref2va_pruned_w4a8_mixed.safetensors |
ComfyUI/models/diffusion_models |
11.8 GB | Withdrawn upstream. Its replacement is Minimax-h3_Singularity_ref2va_v1.3_Pruned_w4a8.safetensors, picked in the loader in its place |
Apache-2.0 |
minimax_h3_video_vae_fp16.safetensors |
ComfyUI/models/vae |
5.2 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_video_vae_int8_convrot.safetensors |
ComfyUI/models/vae |
2.8 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
minimax_h3_audio_vae_fp32.safetensors |
ComfyUI/models/vae |
605 MB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
ComfyUI/models/text_encoders |
15.7 GB | Comfy-Org/MiniMax-H3 | MiniMax H3 Community License |
qwen3vl_8b_fp8_scaled.safetensors |
ComfyUI/models/text_encoders |
10.6 GB | Comfy-Org/Qwen3-VL | Apache-2.0 |
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors |
ComfyUI/models/loras/H3 |
2.0 GB | lightx2v/Minimax-h3-Turbo | Apache-2.0 |
minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors |
ComfyUI/models/loras/H3 |
2.0 GB | lightx2v/Minimax-h3-Turbo | Apache-2.0 |
dmad_minimax_h3_4step_full_critic_comfyui.safetensors |
ComfyUI/models/loras/H3 |
2.0 GB | ZhengmingYu/DMAD | MiniMax H3 Community License |
minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16.safetensors |
ComfyUI/models/loras/H3 |
367 MB | Minimax H3 Turbo LoRas on Civitai, version DMAD_4step_full |
On the Civitai page |
taomate_h3_3step_comfy.safetensors |
ComfyUI/models/loras/H3 |
2.5 GB | Robert1212star/TaoMate-H3-3Step-ComfyUI | On the repository |
Graphs: affine-krea2.json, cns-krea2.json.
| File | Goes in | Size | Download | Licence |
|---|---|---|---|---|
krea2_raw_bf16.safetensors |
ComfyUI/models/diffusion_models |
26.3 GB | Comfy-Org/Krea-2 | Krea 2 Community License |
krea2_raw_fp8_scaled.safetensors |
ComfyUI/models/diffusion_models |
13.1 GB | Comfy-Org/Krea-2 | Krea 2 Community License |
krea2_turbo_fp8_scaled.safetensors |
ComfyUI/models/diffusion_models |
13.1 GB | Comfy-Org/Krea-2 | Krea 2 Community License |
qwen3vl_4b_fp8_scaled.safetensors |
ComfyUI/models/text_encoders |
5.2 GB | Comfy-Org/Krea-2 | Krea 2 Community License |
qwen_image_vae.safetensors |
ComfyUI/models/vae |
254 MB | Comfy-Org/Krea-2 | Krea 2 Community License |
wan_2.1_vae.safetensors |
ComfyUI/models/vae |
254 MB | Comfy-Org/Wan_2.1_ComfyUI_repackaged | Apache-2.0 |
krea2_darkbrush.safetensors |
ComfyUI/models/loras |
469 MB | Comfy-Org/Krea-2 | Krea 2 Community License |
krea2_raw_to_turbo_r256_comfy.safetensors |
ComfyUI/models/loras/Krea2 |
1.7 GB | Krea 2 Turbo LoRA 256dim on Civitai | On the Civitai page |
Graph: cns-klein9b.json. Hugging Face asks for a signed-in account that has accepted the terms
before it serves flux-2-klein-9b-fp8.safetensors.
| File | Goes in | Size | Download | Licence |
|---|---|---|---|---|
flux-2-klein-9b-fp8.safetensors |
ComfyUI/models/diffusion_models |
9.4 GB | black-forest-labs/FLUX.2-klein-9b-fp8 | FLUX Non-Commercial License |
qwen_3_8b_fp8mixed.safetensors |
ComfyUI/models/text_encoders |
8.7 GB | Comfy-Org/vae-text-encorder-for-flux-klein-9b | FLUX Non-Commercial License, as the repository is labelled |
flux2-vae.safetensors |
ComfyUI/models/vae |
336 MB | Comfy-Org/flux2-dev | FLUX.1 [dev] Non-Commercial License, as the repository is labelled |
Graph: flux3-klein-distill-lora.json. Hugging Face asks for a signed-in account that has
accepted the terms before it serves the black-forest-labs files.
| File | Goes in | Size | Download | Licence |
|---|---|---|---|---|
flux3_image_klein_bf16.safetensors |
ComfyUI/models/diffusion_models |
11.1 GB | black-forest-labs/flux-3-image-klein-comfy | FLUX Kommunity License |
flux3_image_klein_sd_bf16.safetensors |
ComfyUI/models/diffusion_models |
11.1 GB | black-forest-labs/flux-3-image-klein-comfy | FLUX Kommunity License |
flux3_image_klein_distill_r256.safetensors |
ComfyUI/models/loras |
1.3 GB | WAS/flux3_klein | FLUX Kommunity License |
flux3_image_klein_distill_r64.safetensors |
ComfyUI/models/loras |
338 MB | WAS/flux3_klein | FLUX Kommunity License |
flux3_image_klein_distill_r32.safetensors |
ComfyUI/models/loras |
169 MB | WAS/flux3_klein | FLUX Kommunity License |
flux3_image_klein_distill_r16.safetensors |
ComfyUI/models/loras |
85 MB | WAS/flux3_klein | FLUX Kommunity License |
qwen3vl_4b_int8_convrot.safetensors |
ComfyUI/models/text_encoders |
4.9 GB | black-forest-labs/flux-3-image-klein-comfy | Apache-2.0, Qwen3-VL-4B-Instruct |
flux2-vae.safetensors |
ComfyUI/models/vae |
336 MB | black-forest-labs/flux-3-image-klein-comfy | FLUX Kommunity License, as the repository is labelled |
The four LoRAs are the distilled checkpoint minus the base, one per rank; the higher the rank,
the closer the base comes to the distilled picture at CFG 1. Flux 3 Flow Map Patch reads the
two ref_time_in weights from flux3_image_klein_sd_bf16.safetensors or
flux3_image_klein_sd_int8_convrot.safetensors, so one distilled file stays in
ComfyUI/models/diffusion_models beside the base.
| What you see | What it means |
|---|---|
| A fine stripe every 4 pixels down the rows, and a blue cast | The LoRA is on the base without Flux 3 Flow Map Patch. Put the patch between the LoRA and the sampler |
holds no ref_time_in weights |
The checkpoint menu names a base file. Pick a file with _sd_ in its name |
takes a FLUX 3 Image model |
The model wired in is another architecture. Wire the Load Diffusion Model on flux3_image_klein_bf16.safetensors |
| File | Goes in | Size | Download | Licence | Graphs |
|---|---|---|---|---|---|
autismmixSDXL_autismmixConfetti.safetensors |
ComfyUI/models/checkpoints |
6.9 GB | AutismMix SDXL on Civitai, version AutismMix_confetti |
On the Civitai page | fast-samplers-and-savers.json, hdr.json |
qwen_3_4b.safetensors |
ComfyUI/models/text_encoders |
8.0 GB | Comfy-Org/z_image_turbo | Apache-2.0 | fast-generate-text.json |
4xNomosWebPhoto_RealPLKSR.safetensors |
ComfyUI/models/upscale_models |
30 MB | Phips/4xNomosWebPhoto_RealPLKSR | CC-BY-4.0 | vtc_temporal_upscale.json |
skin_detail_slider_krea2_ZT-V1.safetensors |
ComfyUI/models/loras/Krea2 |
1.2 MB | Skin Detail Slider on Civitai, version V1 |
On the Civitai page | power-lora-stack.json |
day_&_night_slider_0ts-krea2-v1.safetensors |
ComfyUI/models/loras/Krea2 |
1.2 MB | Day & Night Slider on Civitai, version V1 |
On the Civitai page | power-lora-stack.json |
| What you see | What it means |
|---|---|
The menu reads put a checkpoint in models/EMA-VFI |
Nothing on disk and features.network off. Turn it on, or place a file and restart |
The menu reads put a checkpoint in models/optical_flow |
Nothing on disk and features.network off. Turn it on, or place a file from the optical flow table and restart |
The Prompt Timeline's Preview reads No MiniMax H3 Preview VAE |
No taeh3 file is in a vae_approx folder ComfyUI reads. Hover Unavailable for the folders searched |
The Prompt Timeline's Preview reads Invalid MiniMax H3 Preview VAE |
The file is not the taeh3.safetensors in the MiniMax H3 table. Replace it; the next run reads it |
The Prompt Timeline's Preview reads MiniMax H3 Preview Failed |
Decoding a step raised. Hover Unavailable for the error; sampling goes on without previews |
| The file is on disk but not in the menu | Press R on the canvas |
| A BiRefNet file is on disk but not in the cutout menu | It is a Lite file or another architecture. The menu lists full size BiRefNet and BEN2 base files only |
is a link to ..., which ComfyUI cannot open |
The model folder or file is a link that does not resolve from ComfyUI. Recreate it with mklink on Windows, or point extra_model_paths.yaml at the real folder |
was not found ... Setting features.network: true |
The run wanted a file you have not got, with fetching off |
could not be fetched |
Fetching is on and the download failed. The message carries the reason and the directory to place it in |
only trained to land halfway between two frames |
The multiplier is above 2 and this is not a _t file |
Could not detect model type on a Kandinsky 6 file |
The file is not a Kandinsky 6 transformer. Pick one from the Kandinsky 6 table, or transformer/diffusion_pytorch_model.safetensors from a Kandinsky-6.0 repository |
This file is not the Kandinsky 6 audio VAE |
Kandinsky 6 Audio VAE Loader is set to another file in models/vae |
| Out of memory on Kandinsky 6 Pro | The _w6a8 file is the smallest. Pro runs on a 24 GB card in either quantised form, streamed through ComfyUI's VRAM management |
ComfyUI quits with Fatal Python error: Aborted in VAE Decode after Kandinsky 6 Pro |
Decode with VAE Decode (Tiled): tile_size 512, overlap 64, temporal_size 64, temporal_overlap 8 |
Sound from a pretrain Kandinsky 6 transformer is off |
The pack decodes every non-distilled transformer at audio scale 0.5302, the released checkpoints' value; the pretrained ones use 0.417 |