BAGEL Image Edit

This is the party trick. "Keep the same person, same clothes, but put her on a modern subway reading a newspaper" - and it actually keeps the person. That's BAGEL Image Edit, and it's the reason the model exploded in the first place: it does the kind of instruction-following edit that GPT-4o could do and that Flux-Kontext made you fight for. By spring 2025 people were already fine-tuning on top of it to beat Kontext on image editing. If you're going to run any part of this pack, it's probably this node.

How it works - the dual-path setup

This is the one BAGEL node that wants two views of the same image, and it's also the one where beginners trip. Look at its inputs: it needs both an image (IMAGE) and a vae_latent (LATENT). The image goes to BAGEL's SigLIP/NaViT vision tower; the latent is the same picture run through the official FLUX VAEEncode, giving the model the pixel-level conditioning. Internally it builds a text-only context and an image-only context, then samples with both your text scale and image scale applied.

The catch, from the author's own tooltips: the image you send must be the same preprocessed image you sent to VAEEncode. Same pixels, same size. The workflow that works is:

1. Load your source image.
2. Run it through an official ImageScale node - 16-aligned dimensions, short side at least 512 px, long side at most 1024 px (that's BAGEL's own preprocessing range).
3. Branch that exact resized image: one copy into VAEEncode (FLUX AE loaded by the official VAELoader), the other into this node's image input, and the VAEEncode latent into vae_latent.

The node validates this for you - if the image and latent describe different sizes, it throws a clear error telling you to connect the same preprocessed image to both. Batch size 1 only; one edit per execution.

Inputs and outputs that matter

• prompt - plain-language instruction. The default ("She boards a modern subway, quietly reading a folded newspaper, wearing the same clothes") is literally the model's demo move.
• cfg_text_scale (default 4) - how strongly it follows the edit instruction.
• cfg_img_scale (default 2) - how strongly it preserves the input image's details. This is the fidelity dial: crank it up when edits drift, ease off when the model refuses to change anything.
• cfg_interval (default 0) - start of the CFG interval, end pinned at 1.0.
• timestep_shift (default 3, allowed up to 10 here) - higher favors layout, lower favors detail.
• num_timesteps (default 50), cfg_renorm_type (default text_channel - the official image-edit setting), seed (0 = unset, positive = reproducible).
• The thinking cluster (show_thinking, max_think_tokens, do_sample, text_temperature) - turn it on to have BAGEL plan the edit in text before sampling.

Outputs: latent (LATENT, decode with official VAEDecode + FLUX AE) and reasoning (STRING, the planning text when thinking is enabled).

Installing and what to expect

Shared pack setup: clone https://github.com/neverbiasu/ComfyUI-BAGEL into custom_nodes, pip install -r requirements.txt, drop the single-file BAGEL .safetensors into models/bagel/ and FLUX ae.safetensors into models/vae/, restart. ComfyUI Manager users: search "ComfyUI-BAGEL". The bagel_image_editing.json example workflow is the easiest way to see the ImageScale → VAEEncode wiring done right.

Expect it to be slow and hungry. The README's own validation ran on an A100 and calls high-VRAM the recommendation; the legacy full-precision era reported 30 GB at 1024×1024. If you're under ~24 GB, get an FP8 or INT8 variant of the model first. The node itself is strict about sizes precisely so you fail with a readable error instead of a mysterious broken image - a small mercy in a pack where the model is the bottleneck.