BAGEL Multi-Image Edit (Deprecated, Echo-4o only)

Let's be honest about what this node is: the "(Deprecated, Echo-4o only)" in its display name is the whole story. This is a leftover from the pack's all-in-one era, kept around so old workflows and old model downloads still load. Before the native refactor, ComfyUI-BAGEL auto-downloaded weights, ran its own private pipeline, and used an internal VAE. BagelMultiImageEdit is from that era, and it's aimed at one specific model: Echo-4o, an enhanced BAGEL variant (the Yejy53/Echo-4o family) that gained true multi-reference image support.

So what did it do? Take two to four reference images and a prompt, and produce an edit that combines elements across them - the classic "put this character in that room" composition trick that single-image edit nodes can't pull off. That capability is why the node still exists: the current native nodes don't do multi-image yet (the pack's roadmap lists Echo-4o under "specialized any-to-any / composition" variants to adapt later).

The inputs

Two reference images are required - ref_image_1 and ref_image_2 - with ref_image_3 and ref_image_4 as optional additions. Then:

• prompt - describes the combined edit; the default asks to "Combine elements from the reference images according to the description."
• cfg_text_scale (default 4) - how strongly the edit follows the prompt.
• cfg_img_scale (default 2) - preservation of reference image details.
• num_timesteps (default 50) - denoising steps.
• seed - 0 for random, positive for reproducible.
• Optional extras: show_thinking (returns the reasoning process), cfg_interval, timestep_shift, cfg_renorm_min, cfg_renorm_type (text_channel default), text_temperature.

Outputs: image (IMAGE, already decoded - note this differs from the native nodes, which return a LATENT for you to decode) and thinking (STRING).

Why it's deprecated, and whether you should care

Three reasons, in order of importance. One: it belongs to the old all-in-one loader, which the pack has been migrating away from in favor of native ComfyUI model loading - no auto-download, standard ComfyUI model files, official FLUX VAE. Two: it only really shines with Echo-4o weights, and the native path doesn't support multi-image at all yet, so this node is the only way to do that task at all. Three: the model layouts it expects are the legacy ones.

When to use it: you have an old Echo-4o download or a pre-refactor workflow you need to keep running. For that, the deprecated install doc is your guide - old model folders go in ComfyUI/models/bagel/ (e.g. the sharded ByteDance-Seed/BAGEL-7B-MoT layout), legacy INT8/NF4 quantization wants bitsandbytes, and DFloat11 support needs the dfloat11 package. If you're starting fresh today, don't start here - use the BAGEL Model Loader with a converted single-file model and the native nodes instead. New users landing on this page should treat it as a compatibility bridge, not a destination.