Anima Model Merge Extended (Experimental)

AnimaModelMergeExtended is the experimental variant of Anima Model Merge (Auto Remap), and it adds exactly one knob: blend_ratio. Everything else - the automatic 28-vs-40 block detection, the remap, merge_ratio, extend_ratio, the manifest dropdown - is identical to the regular node.

The question it answers is about the 12 newly-inserted blocks in Anima-2.9B. The regular merge node can blend the old 28-block model's layers into those new blocks via extend_ratio, but it only ever uses the preceding old layer as the source. The Extended node generalizes that to blend both the preceding ("front") and following ("back") old layers at a ratio you pick:

value on new layer = front layer × blend_ratio + back layer × (1 - blend_ratio)

To be clear about how the two knobs interact: extend_ratio controls the mix between "the 2.9B model's own value" and "the blended old-model value" on the new blocks, while blend_ratio controls how that blended old-model value is itself built from front vs. back. At blend_ratio = 1.0 the back layer's contribution is zero, and the README confirms the output is numerically identical to the regular node. So if you're happy with the regular node, this one gives you nothing new at its default - you reach for it specifically when you want to fiddle with how the old model's influence lands on those 12 blocks.

Inputs and outputs

Two MODELs in, one MODEL out:

• model_1 / model_2 - the models to merge.
• merge_ratio (0.0–1.0, default 0.5) - blend weight for model_1 (1.0 = all model_1, 0.0 = all model_2).
• extend_ratio (0.0–1.0, default 0.0) - how much old-model value goes into the new blocks.
• blend_ratio (0.0–1.0, default 1.0) - front vs. back old layer when extending.
• manifest - block-mapping file, one option shipped.

Output is model, which you'd wire to ModelSave (advanced/model_merging) just like the regular node - no built-in save here either, and ModelSave writes the UNet portion only, defaulting to ComfyUI/output.

Same edge case as the Extended LoRA node: the very first and last inserted blocks have only one neighbor, so blend_ratio is ignored there and the sole available neighbor is used at full strength. Otherwise a blend_ratio favoring a nonexistent side could zero out that block's extension entirely.

Should you use it?

The README is upfront: these Extended nodes exist for experimentation and finer control. Both 40-block-in models merge directly, both 28-block models merge directly, and the mixed case produces a 40-block output with the new layers staying 2.9B-side by default - the same as the regular node. If you don't feel a need to tune the new-block projection, save yourself a search term and use the regular node. If you do want the extra axis, this is where it lives, and blend_ratio = 1.0 gives you a clean baseline to compare against.

Install

Same pack, same path:

cd ComfyUI/custom_nodes
git clone https://github.com/shin131002/ComfyUI-Anima29B-Remap.git

or ComfyUI Manager → "Anima-2.9B Remap", then restart. No extra dependencies or downloads. The repo is frozen in favor of ComfyUI-Anima-Remap (which also handles Anima-3.8B's 52-block layout) - don't install both, the node IDs collide. And if you end up saving a merged model, remember it's a non-commercial derivative under the CircleStone Labs license, even though the images it produces are free to use commercially.