ComfyUI-BAGEL
A ComfyUI custom node package for BAGEL-7B-MoT with native ComfyUI model loading.
Model paths
| Component |
Recommended source |
Put it here |
Loaded by |
| BAGEL main model |
6chan/bagel_comfy single-file .safetensors |
ComfyUI/models/bagel/ |
BAGEL Model Loader |
| FLUX AE / VAE |
ae.safetensors for FLUX |
ComfyUI/models/vae/ |
official VAELoader, VAEEncode, VAEDecode |
| Qwen tokenizer |
bundled in this repository |
no manual install |
BAGEL Model Loader |
| BAGEL model configs |
built into the node; optional metadata can override them |
no manual install |
BAGEL Model Loader |
| Old HF shard layout |
ByteDance-Seed/BAGEL-7B-MoT, DFloat11/BAGEL-7B-MoT-DF11 |
ComfyUI/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
-
Clone this repository into ComfyUI/custom_nodes/ComfyUI-BAGEL.
-
Install the node-specific dependencies:
cd ComfyUI/custom_nodes/ComfyUI-BAGEL
pip install -r requirements.txt
-
Download a BAGEL .safetensors from 6chan/bagel_comfy into ComfyUI/models/bagel/.
-
Put FLUX ae.safetensors into ComfyUI/models/vae/.
-
Restart ComfyUI and load one of the native workflows below.
Existing old-model users
| Current state |
Recommended action |
You already downloaded ByteDance-Seed/BAGEL-7B-MoT |
Convert it with scripts/convert_bagel_model.py, or re-download the converted single-file model. |
You already downloaded DFloat11/BAGEL-7B-MoT-DF11 |
Keep using deprecated workflows for now, or convert/re-download when a converted quantized release is available. |
| You have old all-in-one BAGEL workflows |
Use 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
| Workflow |
File |
Extra nodes |
Notes |
| Text-to-image |
example_workflows/bagel_text_to_image.json |
none |
Native BAGEL latent generation, official VAEDecode. |
| Image editing |
example_workflows/bagel_image_editing.json |
official ImageScale, VAEEncode, VAEDecode |
Resize 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 understanding |
example_workflows/bagel_image_understanding.json |
`ShowText |
pysssssfromcomfyui-custom-scripts, or replace with official Preview as Text` on newer ComfyUI |
| Deprecated text-to-image |
example_workflows/bagel_text_to_image_deprecated.json |
comfyui-custom-scripts |
Old all-in-one loader. |
| Deprecated image editing |
example_workflows/bagel_image_editing_deprecated.json |
comfyui-custom-scripts |
Old all-in-one loader. |
| Deprecated image understanding |
example_workflows/bagel_image_understanding_deprecated.json |
comfyui-custom-scripts |
Old all-in-one loader. |
Model and runtime matrix
| Path |
Model source |
File layout |
Nodes |
VAE |
Status |
| Native BF16 |
6chan/bagel_comfy |
single .safetensors in models/bagel |
BAGEL* native nodes |
official FLUX AE |
recommended |
| Converted local BF16 |
original ByteDance-Seed/BAGEL-7B-MoT converted by script |
single .safetensors in models/bagel |
BAGEL* native nodes |
official FLUX AE |
supported |
| Legacy standard |
original HF shard folder |
folder in models/bagel |
Bagel* (Deprecated) |
internal legacy VAE |
compatibility only |
| Legacy DFloat11 |
DFloat11 HF folder |
folder in models/bagel |
Bagel* (Deprecated) |
internal legacy VAE |
compatibility only |
| Task |
Recommended workflow |
Expected VRAM |
Current validation |
| Text-to-image |
native BF16 + official VAE decode |
A100-class / high-VRAM GPU recommended |
ran on Modal A100; exact benchmark pending |
| Image editing |
native BF16 + official VAE encode/decode |
A100-class / high-VRAM GPU recommended |
ran on Modal A100; exact benchmark pending |
| Image understanding |
native BF16, no VAE |
lower than generation/editing, still needs BAGEL loaded |
workflow prepared; run on remote GPU |
| Legacy DFloat11 generation |
deprecated workflows |
21.76 GB reported for 1024x1024 |
old README reported 154.39 s on RTX 4090 |
| Legacy standard generation |
deprecated workflows |
30.07 GB reported for 1024x1024 |
old 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 family |
Example models in collection |
Input / output shape |
Planned ComfyUI support |
| Base BAGEL any-to-any |
ByteDance-Seed/BAGEL-7B-MoT, 6chan/bagel_comfy |
text, image, latent -> text/image |
current native nodes |
| Image editing / NHR editing |
iitolstykh/Bagel-NHR-Edit, Bagel-NHR-Edit-V2 |
image + edit prompt -> image |
first try native image-edit adapter; add a dedicated edit node if prompt/image order differs |
| Reasoning / VQA variants |
multimodal-reasoning-lab/Bagel-Zebra-CoT, sensenova/SenseNova-SI-1.1-BAGEL-7B-MoT |
image + text -> text |
adapt BAGEL Image Understanding; add reasoning-specific controls if needed |
| Text-to-image variants |
Wayne-King/SRUM_BAGEL_7B_MoT, Ryann829/Scone, LLM-Drop/*GEN*, Yanran21/UniGenDet |
text -> image |
adapt BAGEL Text to Image; add model-specific generation options only when required |
| SIGMA variants |
SIGMA checkpoints in the collection and related releases |
style/subject or structured conditioning -> image |
add a SIGMA-specific loader/conditioning node after its checkpoint metadata and reference workflow are verified |
| Quantized formats |
DFloat11/*, FP8, INT8, GGUF, AutoRound INT4 |
same tasks, different weight format/runtime |
separate loader/converter path; do not mix into the BF16 loader unless the state dict is compatible |
| Specialized any-to-any / composition |
ThinkMorph, Uni-Edit, UniCorn, ConsistCompose, Echo-4o, SenseNova-Vision |
may add multi-image, identity, composition, or agentic conditioning |
inspect model card + sample code first; add new nodes when graph inputs differ from the base BAGEL tasks |
| Support stage |
Acceptance gate |
| Catalog |
Add model to a comparison table with source, license, task, file format, expected VRAM, and required nodes. |
| Load |
Converted or downloaded model appears in the correct ComfyUI model folder and loads without auto-downloading pipeline code. |
| Smoke test |
One minimal workflow runs on a remote GPU for the model’s primary task. |
| Workflow |
Add or update an example workflow with official ComfyUI nodes where possible. |
| Benchmark |
Record 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.
评论