MugenMaluMix v4.2
Sharing the merge recipe for the Pony Diffusion merge checkpoint MugenMaluMix mix version 4.
This version is inspired by the DARE merging described by the paper Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch
The intended use of this merge checkpoint is to produce Game Art, Anime, and Character subjects in a Fantasy style.
New in this version:
Used DARE merging through ComfyUI with gradient masking to create the base model
Merged additional new finetune checkpoints:
Applied new styles and concepts though the inclusion of new LoRAs
Continued use of a hierachical merge approach to asses the priority & weights of each LoRA category (Styles, Concepts, Characters)
Improved responsiveness to and compatibility with "masterpeice, best quality, very aesthetic, absurdres" quality tagging
Reduced the oily/shiny appearance of skin
Better high dynamic range, color saturation and contrast of base outputs
Quality goals remain to prioritise better anatomy (hands, feet particularly) and maintain the base models responsiveness to booru tagging.
📔Resource Overview
🚩Checkpoints
🖌LoRAs by author motimalu
Grouped by homogulous traits for DARE LoRA merging
Game Art:
Akihiko Yoshida Style (NieR, Bravely Default, FF Tactics) [Pony]
Star Ocean: The Second Story R - Yukihiro Kajimoto Style [Pony]
Artists:
Monmusu:
Fluffy Tail
Frieren:
Mushoku:
🖌Community LoRAs
"RGB Gemstone" group, these give great improvement to color saturation, contrast, and fine details on jewelery, eyes reflections and lighting:
🧙♂️Merge setings
Merge tool: ComfyUI
Merge Nodes from:
Merge Strategy:
1: Hierachical DARE Checkpoint merge with gradient masking
2: DARE Merge of LoRA stacks sorted into similar categoties are created
3: Weights for each LoRA are determined though testing before baking into the final model
ℹ️ComfyUI workflows are embedded in the images
To test out/reuse the merge recipes ComfyUI and the above DARE nodes are required.
ComfyUI JSON for the DARE base model merge and DARE LoRA finetuning have also been added to the Model Training Files.
⚠️These are quite messy, and will require significant hardware to run without modification - I used a machine with 64gb of system ram.
Step 1. Sequential DARE merge of base models
AutismMix SDXL Confetti merged with 4th tail (Hentai model) v.0.4.0
Hassaku XL (Hentai model) v1.1 merged with Omega Pony XL anime style
The intermediate models are then combined via a final DARE merge:
Step 2. DARE merge of LoRA groups
Game Art:
Artists:
Monmusu:
Fluffy Tail:
RGB Gemstones:
Frieren:
Mushoku:
Step 3. LoRA merge to checkpoint
X/Y Comparisons of the DARE LoRA's effect to determine a final weights to bake into the checkpoint:
With a bit more testing of how these work in tandem, I decided on these weights:
rgb_gemstones: 0.42
mushoku: 0.40
gameart: 0.25
monmusu: 0.25
fluffytail: 0.25
frieren: 0.20
artists: 0.15
With the desired weights determined, the LoRA patch to the checkpoint is saved as the LoRA finetune model: mugenmalumixSDXL_v42
🧪Results:
X/Y checkpoint comparison plots of some results for prompts that emphasize characteristics of this merge (on catbox to allow for the large file sizes):
That's it!
This version took quite a while including a few re-makes and research to familiarise with comfy-ui DARE merging.
As always, please feel free to share your thoughts/feedback in the comments below. ´ ˘ `