neverbiasu/ComfyUI-BAGEL

A ComfyUI extention for BAGEL(Unified Model for Multimodal Understanding and Generation)

187

stars

34

commits

Python

primary language

Jul 31, 2026

updated

comfyui
comfyui-custom-node
comfyui-nodes
custom-nodes
Browse cluster: ComfyUI Custom Nodes and Extensions

README

ComfyUI-BAGEL

A ComfyUI custom node package for BAGEL-7B-MoT with native ComfyUI model loading.

BAGEL

Model paths

ComponentRecommended sourcePut it hereLoaded by
BAGEL main model6chan/bagel_comfy single-file .safetensorsComfyUI/models/bagel/BAGEL Model Loader
FLUX AE / VAEae.safetensors for FLUXComfyUI/models/vae/official VAELoader, VAEEncode, VAEDecode
Qwen tokenizerbundled in this repositoryno manual installBAGEL Model Loader
BAGEL model configsbuilt into the node; optional metadata can override themno manual installBAGEL Model Loader
Old HF shard layoutByteDance-Seed/BAGEL-7B-MoT, DFloat11/BAGEL-7B-MoT-DF11ComfyUI/models/bagel/deprecated legacy nodes

Recommended: download the single-file model from 6chan/bagel_comfy. If you already have the original BAGEL checkpoint, either convert it with scripts/convert_bagel_model.py or re-download the converted file. The converter embeds and validates the comfyui_bagel metadata header, including the model configs needed by the native loader. If the Hugging Face repository also contains config files, treat them as conversion/audit references. The native ComfyUI loader does not require users to copy config files into models/bagel.

Install

  1. Clone this repository into ComfyUI/custom_nodes/ComfyUI-BAGEL.

  2. Install the node-specific dependencies:

    cd ComfyUI/custom_nodes/ComfyUI-BAGEL
    pip install -r requirements.txt
    
  3. Download a BAGEL .safetensors from 6chan/bagel_comfy into ComfyUI/models/bagel/.

  4. Put FLUX ae.safetensors into ComfyUI/models/vae/.

  5. Restart ComfyUI and load one of the native workflows below.

Existing old-model users

Current stateRecommended action
You already downloaded ByteDance-Seed/BAGEL-7B-MoTConvert it with scripts/convert_bagel_model.py, or re-download the converted single-file model.
You already downloaded DFloat11/BAGEL-7B-MoT-DF11Keep using deprecated workflows for now, or convert/re-download when a converted quantized release is available.
You have old all-in-one BAGEL workflowsUse the _deprecated workflow files and deprecated nodes, then migrate to native workflows.

Legacy auto-download and all-in-one loader instructions were moved to docs/deprecated-installation.md.

Workflows

BAGEL text-to-image workflow

BAGEL image-editing workflow

BAGEL image-understanding workflow

WorkflowFileExtra nodesNotes
Text-to-imageexample_workflows/bagel_text_to_image.jsonnoneNative BAGEL latent generation, official VAEDecode.
Image editingexample_workflows/bagel_image_editing.jsonofficial ImageScale, VAEEncode, VAEDecodeResize the source image before both VAEEncode and BAGEL Image Edit; use 16-aligned dimensions, with a 512–1024px range matching the original BAGEL preprocessing.
Image understandingexample_workflows/bagel_image_understanding.json`ShowTextpysssssfromcomfyui-custom-scripts, or replace with official Preview as Text` on newer ComfyUI
Deprecated text-to-imageexample_workflows/bagel_text_to_image_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.
Deprecated image editingexample_workflows/bagel_image_editing_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.
Deprecated image understandingexample_workflows/bagel_image_understanding_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.

Model and runtime matrix

PathModel sourceFile layoutNodesVAEStatus
Native BF166chan/bagel_comfysingle .safetensors in models/bagelBAGEL* native nodesofficial FLUX AErecommended
Converted local BF16original ByteDance-Seed/BAGEL-7B-MoT converted by scriptsingle .safetensors in models/bagelBAGEL* native nodesofficial FLUX AEsupported
Legacy standardoriginal HF shard folderfolder in models/bagelBagel* (Deprecated)internal legacy VAEcompatibility only
Legacy DFloat11DFloat11 HF folderfolder in models/bagelBagel* (Deprecated)internal legacy VAEcompatibility only
TaskRecommended workflowExpected VRAMCurrent validation
Text-to-imagenative BF16 + official VAE decodeA100-class / high-VRAM GPU recommendedran on Modal A100; exact benchmark pending
Image editingnative BF16 + official VAE encode/decodeA100-class / high-VRAM GPU recommendedran on Modal A100; exact benchmark pending
Image understandingnative BF16, no VAElower than generation/editing, still needs BAGEL loadedworkflow prepared; run on remote GPU
Legacy DFloat11 generationdeprecated workflows21.76 GB reported for 1024x1024old README reported 154.39 s on RTX 4090
Legacy standard generationdeprecated workflows30.07 GB reported for 1024x1024old README reported 482.95 s on RTX 4090

TODO: model variant support

Future BAGEL variants will be tracked from the 6chan/bagel collection. The support path depends on each model's file format, modality inputs, and inference logic.

Variant familyExample models in collectionInput / output shapePlanned ComfyUI support
Base BAGEL any-to-anyByteDance-Seed/BAGEL-7B-MoT, 6chan/bagel_comfytext, image, latent -> text/imagecurrent native nodes
Image editing / NHR editingiitolstykh/Bagel-NHR-Edit, Bagel-NHR-Edit-V2image + edit prompt -> imagefirst try native image-edit adapter; add a dedicated edit node if prompt/image order differs
Reasoning / VQA variantsmultimodal-reasoning-lab/Bagel-Zebra-CoT, sensenova/SenseNova-SI-1.1-BAGEL-7B-MoTimage + text -> textadapt BAGEL Image Understanding; add reasoning-specific controls if needed
Text-to-image variantsWayne-King/SRUM_BAGEL_7B_MoT, Ryann829/Scone, LLM-Drop/*GEN*, Yanran21/UniGenDettext -> imageadapt BAGEL Text to Image; add model-specific generation options only when required
SIGMA variantsSIGMA checkpoints in the collection and related releasesstyle/subject or structured conditioning -> imageadd a SIGMA-specific loader/conditioning node after its checkpoint metadata and reference workflow are verified
Quantized formatsDFloat11/*, FP8, INT8, GGUF, AutoRound INT4same tasks, different weight format/runtimeseparate loader/converter path; do not mix into the BF16 loader unless the state dict is compatible
Specialized any-to-any / compositionThinkMorph, Uni-Edit, UniCorn, ConsistCompose, Echo-4o, SenseNova-Visionmay add multi-image, identity, composition, or agentic conditioninginspect model card + sample code first; add new nodes when graph inputs differ from the base BAGEL tasks
Support stageAcceptance gate
CatalogAdd model to a comparison table with source, license, task, file format, expected VRAM, and required nodes.
LoadConverted or downloaded model appears in the correct ComfyUI model folder and loads without auto-downloading pipeline code.
Smoke testOne minimal workflow runs on a remote GPU for the model's primary task.
WorkflowAdd or update an example workflow with official ComfyUI nodes where possible.
BenchmarkRecord GPU, VRAM peak, image size, steps, and wall-clock time.

Contribution

Contributions are welcome. Please open an issue for compatibility problems, variant support requests, or proposed changes before submitting a pull request. For code changes, include the affected workflow and the ComfyUI version used for validation when possible.

FAQ

Which model should I download?

Use the single-file converted checkpoints from 6chan/bagel_comfy. The native loader does not require users to copy the original BAGEL config files into models/bagel.

Which VAE should I use?

Load the FLUX autoencoder with the official VAELoader, then connect it to VAEEncode/VAEDecode. The native BAGEL nodes do not load a private pipeline VAE.

What image sizes are supported for editing?

Resize the source image before both VAEEncode and BAGEL Image Edit. Use 16-aligned dimensions, normally with the short side at least 512 px and the long side at most 1024 px.

Can I still use an original BAGEL or DFloat11 model?

Yes, but use the deprecated workflows and nodes. For native workflows, convert the checkpoint or download the matching single-file release.

License

This project is licensed under the Apache 2.0 License. Please refer to the official license terms for the use of the BAGEL model.

Contributors

neverbiasu

33 commits

wildminder

1 commits

neverbiasu/ComfyUI-BAGEL

A ComfyUI extention for BAGEL(Unified Model for Multimodal Understanding and Generation)

187

stars

34

commits

Python

primary language

Jul 31, 2026

updated

comfyui
comfyui-custom-node
comfyui-nodes
custom-nodes
Browse cluster: ComfyUI Custom Nodes and Extensions

README

ComfyUI-BAGEL

A ComfyUI custom node package for BAGEL-7B-MoT with native ComfyUI model loading.

BAGEL

Model paths

ComponentRecommended sourcePut it hereLoaded by
BAGEL main model6chan/bagel_comfy single-file .safetensorsComfyUI/models/bagel/BAGEL Model Loader
FLUX AE / VAEae.safetensors for FLUXComfyUI/models/vae/official VAELoader, VAEEncode, VAEDecode
Qwen tokenizerbundled in this repositoryno manual installBAGEL Model Loader
BAGEL model configsbuilt into the node; optional metadata can override themno manual installBAGEL Model Loader
Old HF shard layoutByteDance-Seed/BAGEL-7B-MoT, DFloat11/BAGEL-7B-MoT-DF11ComfyUI/models/bagel/deprecated legacy nodes

Recommended: download the single-file model from 6chan/bagel_comfy. If you already have the original BAGEL checkpoint, either convert it with scripts/convert_bagel_model.py or re-download the converted file. The converter embeds and validates the comfyui_bagel metadata header, including the model configs needed by the native loader. If the Hugging Face repository also contains config files, treat them as conversion/audit references. The native ComfyUI loader does not require users to copy config files into models/bagel.

Install

  1. Clone this repository into ComfyUI/custom_nodes/ComfyUI-BAGEL.

  2. Install the node-specific dependencies:

    cd ComfyUI/custom_nodes/ComfyUI-BAGEL
    pip install -r requirements.txt
    
  3. Download a BAGEL .safetensors from 6chan/bagel_comfy into ComfyUI/models/bagel/.

  4. Put FLUX ae.safetensors into ComfyUI/models/vae/.

  5. Restart ComfyUI and load one of the native workflows below.

Existing old-model users

Current stateRecommended action
You already downloaded ByteDance-Seed/BAGEL-7B-MoTConvert it with scripts/convert_bagel_model.py, or re-download the converted single-file model.
You already downloaded DFloat11/BAGEL-7B-MoT-DF11Keep using deprecated workflows for now, or convert/re-download when a converted quantized release is available.
You have old all-in-one BAGEL workflowsUse the _deprecated workflow files and deprecated nodes, then migrate to native workflows.

Legacy auto-download and all-in-one loader instructions were moved to docs/deprecated-installation.md.

Workflows

BAGEL text-to-image workflow

BAGEL image-editing workflow

BAGEL image-understanding workflow

WorkflowFileExtra nodesNotes
Text-to-imageexample_workflows/bagel_text_to_image.jsonnoneNative BAGEL latent generation, official VAEDecode.
Image editingexample_workflows/bagel_image_editing.jsonofficial ImageScale, VAEEncode, VAEDecodeResize the source image before both VAEEncode and BAGEL Image Edit; use 16-aligned dimensions, with a 512–1024px range matching the original BAGEL preprocessing.
Image understandingexample_workflows/bagel_image_understanding.json`ShowTextpysssssfromcomfyui-custom-scripts, or replace with official Preview as Text` on newer ComfyUI
Deprecated text-to-imageexample_workflows/bagel_text_to_image_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.
Deprecated image editingexample_workflows/bagel_image_editing_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.
Deprecated image understandingexample_workflows/bagel_image_understanding_deprecated.jsoncomfyui-custom-scriptsOld all-in-one loader.

Model and runtime matrix

PathModel sourceFile layoutNodesVAEStatus
Native BF166chan/bagel_comfysingle .safetensors in models/bagelBAGEL* native nodesofficial FLUX AErecommended
Converted local BF16original ByteDance-Seed/BAGEL-7B-MoT converted by scriptsingle .safetensors in models/bagelBAGEL* native nodesofficial FLUX AEsupported
Legacy standardoriginal HF shard folderfolder in models/bagelBagel* (Deprecated)internal legacy VAEcompatibility only
Legacy DFloat11DFloat11 HF folderfolder in models/bagelBagel* (Deprecated)internal legacy VAEcompatibility only
TaskRecommended workflowExpected VRAMCurrent validation
Text-to-imagenative BF16 + official VAE decodeA100-class / high-VRAM GPU recommendedran on Modal A100; exact benchmark pending
Image editingnative BF16 + official VAE encode/decodeA100-class / high-VRAM GPU recommendedran on Modal A100; exact benchmark pending
Image understandingnative BF16, no VAElower than generation/editing, still needs BAGEL loadedworkflow prepared; run on remote GPU
Legacy DFloat11 generationdeprecated workflows21.76 GB reported for 1024x1024old README reported 154.39 s on RTX 4090
Legacy standard generationdeprecated workflows30.07 GB reported for 1024x1024old README reported 482.95 s on RTX 4090

TODO: model variant support

Future BAGEL variants will be tracked from the 6chan/bagel collection. The support path depends on each model's file format, modality inputs, and inference logic.

Variant familyExample models in collectionInput / output shapePlanned ComfyUI support
Base BAGEL any-to-anyByteDance-Seed/BAGEL-7B-MoT, 6chan/bagel_comfytext, image, latent -> text/imagecurrent native nodes
Image editing / NHR editingiitolstykh/Bagel-NHR-Edit, Bagel-NHR-Edit-V2image + edit prompt -> imagefirst try native image-edit adapter; add a dedicated edit node if prompt/image order differs
Reasoning / VQA variantsmultimodal-reasoning-lab/Bagel-Zebra-CoT, sensenova/SenseNova-SI-1.1-BAGEL-7B-MoTimage + text -> textadapt BAGEL Image Understanding; add reasoning-specific controls if needed
Text-to-image variantsWayne-King/SRUM_BAGEL_7B_MoT, Ryann829/Scone, LLM-Drop/*GEN*, Yanran21/UniGenDettext -> imageadapt BAGEL Text to Image; add model-specific generation options only when required
SIGMA variantsSIGMA checkpoints in the collection and related releasesstyle/subject or structured conditioning -> imageadd a SIGMA-specific loader/conditioning node after its checkpoint metadata and reference workflow are verified
Quantized formatsDFloat11/*, FP8, INT8, GGUF, AutoRound INT4same tasks, different weight format/runtimeseparate loader/converter path; do not mix into the BF16 loader unless the state dict is compatible
Specialized any-to-any / compositionThinkMorph, Uni-Edit, UniCorn, ConsistCompose, Echo-4o, SenseNova-Visionmay add multi-image, identity, composition, or agentic conditioninginspect model card + sample code first; add new nodes when graph inputs differ from the base BAGEL tasks
Support stageAcceptance gate
CatalogAdd model to a comparison table with source, license, task, file format, expected VRAM, and required nodes.
LoadConverted or downloaded model appears in the correct ComfyUI model folder and loads without auto-downloading pipeline code.
Smoke testOne minimal workflow runs on a remote GPU for the model's primary task.
WorkflowAdd or update an example workflow with official ComfyUI nodes where possible.
BenchmarkRecord GPU, VRAM peak, image size, steps, and wall-clock time.

Contribution

Contributions are welcome. Please open an issue for compatibility problems, variant support requests, or proposed changes before submitting a pull request. For code changes, include the affected workflow and the ComfyUI version used for validation when possible.

FAQ

Which model should I download?

Use the single-file converted checkpoints from 6chan/bagel_comfy. The native loader does not require users to copy the original BAGEL config files into models/bagel.

Which VAE should I use?

Load the FLUX autoencoder with the official VAELoader, then connect it to VAEEncode/VAEDecode. The native BAGEL nodes do not load a private pipeline VAE.

What image sizes are supported for editing?

Resize the source image before both VAEEncode and BAGEL Image Edit. Use 16-aligned dimensions, normally with the short side at least 512 px and the long side at most 1024 px.

Can I still use an original BAGEL or DFloat11 model?

Yes, but use the deprecated workflows and nodes. For native workflows, convert the checkpoint or download the matching single-file release.

License

This project is licensed under the Apache 2.0 License. Please refer to the official license terms for the use of the BAGEL model.

Contributors

neverbiasu

33 commits

wildminder

1 commits

Languages

Python

99.6%