Arthemy Merge - ComfyUI — Anima v1.0

Arthemy Merge - ComfyUI is for shaping Anima models and their CLIP encoders in ComfyUI. It can apply block-level LoRA edits, merge or tune model and CLIP blocks, repair weight drift, and save modified MODEL and CLIP as .safetensors files. It uses a Base UNET, CLIP, VAE, LoRA, a�

Block merger nodes slice the UNET into 7 groups and CLIP into 5 sub-modules, so you can choose which parts come from Model 1 or Model 2.

Chaos merger and tuner nodes assign weights randomly using a Seed when merging two models or tuning one; locking the Seed makes the result reproducible. The groups can also be concatenated to shape the model and CLIP.

Standard loaders supply the Base UNET, CLIP, and VAE; KSampler uses a test prompt and seed to produce an image for inspection.

After testing, connect the modified MODEL and CLIP to ArthemyAnimaModelSaver and CLIPSaver to bake the weights into .safetensors files.

For controlled comparisons, place loaders at the beginning, savers at the end, keep the Seed fixed, and use a specific, complex test prompt that can reveal elements appearing or disappearing.

Vorschlag · nicht geprüft

The workflow names these model files: 3D_Anime_Style.safetensors, anima_baseV10.safetensors, Arthemy_Comics_Anima_v2.safetensors, and Arthemy_Qwen3_6b.safetensors.

It also names LessAnima_epoch_10.safetensors, qwen_3_06b_base.safetensors, qwen_image_vae.safetensors, and untitled-step00002400.safetensors.

Named nodes include: ArthemyAnimaLoraBlockLoader, ArthemyAnimaModelBlockMerger, ArthemyAnimaCLIPBlockMerger, ArthemyAnimaModelTuner, ArthemyAnimaCLIPTuner, ArthemyAnimaModelChaosBlockMerger, ArthemyAnimaCLIPChaosBlockMerger, ArthemyAnimaModelChaosBlockTuner, ArthemyAnimaCLIPChaos

ArthemyAnimaLoraBlockLoader applies LoRA weights only to selected semantic blocks with a 0.0–1.0 slider; negative values such as -0.5 are also supported.

ModelTuner and CLIPTuner scale selected block groups: Soft Value uses a smoothed curve, while Real Value performs direct multiplication. ModelRestorer and CLIPRestorer use global_variance_limit to reduce extreme weight spikes and global_offset_fix to move weight distributions

Block merger sliders can be set above 1.0 or below 0.0 to over-amplify a trait.

Treat labels such as Subject_Identity and Art_Style_Medium as nicknames for now; their exact behavior may be updated. Also verify that qwen_image_vae.safetensors is the intended VAE file for your setup.

Vorschlag · nicht geprüft