Update README.md

README.md

@@ -3,7 +3,7 @@ license: other
 license_name: fair-ai-public-license-1.0-sd
 license_link: https://freedevproject.org/faipl-1.0-sd/
 base_model:
-- CabalResearch/NoobAI-Flux2VAE-RectifiedFlow
+- CabalResearch/NoobAI-Flux2VAE-RectifiedFlow-0.3
 library_name: diffusers
 ---
 
@@ -12,16 +12,34 @@ library_name: diffusers
 
 ## Model Description
 
-Mugen is a continuation of the [NoobAI Flux2 v0.3](https://huggingface.co/CabalResearch/NoobAI-Flux2VAE-RectifiedFlow-0.3), renamed to signify a substantial divergence from the original NoobAI models.
+Mugen is a continuation of our SDXL to Flux 2 VAE conversion, renamed to signify a substantial divergence from the original NoobAI models.
 
-It has been trained for 7 additional epochs.
+It has been trained for 7 additional epochs, totalling under **8000$** for a full latent space conversion, while preserving and improving upon model anime knowledge.
+
+In particular, we have paid attention to character performance in this iteration, and developed in-house approach for benchmarking their performance, about which you can read below.
+
+Overall, model performs particularly well with textures and patterns that were previously simply impossible due to SDXL VAE. We prioritized keeping our training as standard-friendly as possible, so local community can easily train on it like on a new Base Model, which it practically is.
 
 
 - **Developed by:** Cabal Research (Bluvoll, Anzhc)
 - **Funded by:** Community
 - **License:** [fair-ai-public-license-1.0-sd](https://freedevproject.org/faipl-1.0-sd/)
-- **Finetuned from model:** [NoobAI Flux2 VAE experiment](https://huggingface.co/CabalResearch/NoobAI-Flux2VAE-RectifiedFlow)
+- **Resumed from:** [NoobAI Flux2 VAE v0.3](https://huggingface.co/CabalResearch/NoobAI-Flux2VAE-RectifiedFlow)
+
+### Character Knowledge Benchmark
+
+![series_dashboard](https://cdn-uploads.huggingface.co/production/uploads/633b43d29fe04b13f46c8988/trBLD3zcXDsMA1Iwuc4fj.png)
+
+
+![unified_characters_standalone](https://cdn-uploads.huggingface.co/production/uploads/633b43d29fe04b13f46c8988/9RIxkWpmT9j_z2uGOWIXa.png)
+
+This benchmark measures character similarity across 1815 characters in this iteration of it.
+
+We utilize reference(non-generated) set of images, and measure character features against ai-generated data - this is the similarity score. Our custom in-house model for character discrimination trained on ~1.2kk images is used.
 
+Due to compute constraint, we selected only single model to compare against - not yet released latest version of [Chenkin](https://huggingface.co/ChenkinNoob/ChenkinNoob-XL-V0.3-BETA) model, which is the most trained SDXL-based anime model currently.
+
+Fuure benchmark iterations might include different arches and more models.
 
 ## Bias and Limitations
 
@@ -29,7 +47,6 @@ General data biases from Danbooru might apply.
 
 Flux 2 VAE seem to have brown bias overall, which can be alleviated by adding `sepia` or `brown theme` to negative.
 
-
 ## Model Output Examples
 
 
@@ -41,13 +58,14 @@ Flux 2 VAE seem to have brown bias overall, which can be alleviated by adding `s
 
 #### Comfy
 
-![image](https://cdn-uploads.huggingface.co/production/uploads/633b43d29fe04b13f46c8988/urIE5XGaUpjkaXutaiebs.png)
+![изображение](https://cdn-uploads.huggingface.co/production/uploads/633b43d29fe04b13f46c8988/668lIifTlZ8sYzWlzLVc0.png)
+
+(Workflow is available alongside model in repo)  
 
-(Workflow is available alongside model in repo)
 We will provide a Node, and hope it will be adapted natively in main repo eventually:  
 **https://github.com/Anzhc/SDXL-Flux2VAE-ComfyUI-Node**
 
-Seems like you don't need to use the node itself, patch is applied without it.
+Just install it, and it will patch the model config, no mode changes required.
 
 
 Apparently works in [SwarmUI](https://github.com/mcmonkeyprojects/SwarmUI) as is.
@@ -57,12 +75,15 @@ Same as your normal inference, but with addition of SD3 sampling node, as this m
 Recommended Parameters:  
 **Sampler**: Euler A, Euler, DPM++ SDE, etc.  
 **Steps**: 20-28  
-**CFG**: 4-6  
+**CFG**: 4-7  
 **Shift**: 8-12  
 **Schedule**: Normal/Simple/SGM Uniform  
 **Positive Quality Tags**: `masterpiece, best quality`  
 **Negative Tags**: `worst quality, normal quality, bad anatomy, sepia`   
 
+**Alternative Extended Negative**: `(worst quality:1.1), normal quality, (bad anatomy:1.1), (blurry:1.1), watermark, sepia, (adversarial noise:1.1), jpeg artifacts`  
+(Some of our testers pointed out that they prefer longer negative)
+
 
 #### A1111 WebUI
 
@@ -91,12 +112,15 @@ Flux2VAE does not currently have an appropriate high quality preview method, ple
 Recommended Parameters:  
 **Sampler**: Euler A Comfy RF, Euler A2, Euler, DPM++ SDE Comfy, etc. **ALL VARIANTS MUST BE RF OR COMFY, IF AVAILABLE. In ComfyUI routing is automatic, but not in the case of WebUI.**  
 **Steps**: 20-28  
-**CFG**: 4-6(or 7-15, if it appears to be weak/bugged)  
+**CFG**: 4-7(or 7-15, if it appears to be weak/bugged)  
 **Shift**: 8-12  
 **Schedule**: Normal/Simple/SGM Uniform  
 **Positive Quality Tags**: `masterpiece, best quality`  
 **Negative Tags**: `worst quality, normal quality, bad anatomy, sepia`  
 
+**Alternative Extended Negative**: `(worst quality:1.1), normal quality, (bad anatomy:1.1), (blurry:1.1), watermark, sepia, (adversarial noise:1.1), jpeg artifacts`  
+(Some of our testers pointed out that they prefer longer negative)
+
 **ADETAILER FIX FOR RF**:
 By default, Adetailer discards Advanced Model Sampling extension, which breaks RF. You need to add AMS to this part of settings:
 
@@ -120,17 +144,10 @@ _builtin_script = (
 
 Or use my fork of Adetailer - https://github.com/Anzhc/aadetailer-reforge
 
-## Training
-
-VAE: [Flux2 VAE](https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main/vae)
-
-**VAE Shift**: 0.0760  
-**VAE Scale**: 0.6043
-
 
 ### LoRA Training
 
-How about we just give config
+You can directly reference config with all parameters: Download
 
 ### Hardware