| Property | Value |
|---|---|
| License | MIT |
| Base Model | black-forest-labs/FLUX.1-dev |
| Research Paper | On Distillation of Guided Diffusion Models |
| Language | English |
What is flux-dev-de-distill?
flux-dev-de-distill is an experimental AI model that aims to reverse the distillation process of guidance from FLUX.1-dev. Created by nyanko7, this model implements true classifier-free guidance (CFG) by removing the original distilled guidance and reconstructing it from scratch.
Implementation Details
The model follows Algorithm 1 from the paper "On Distillation of Guided Diffusion Models," implementing a student model x(zt) that matches the teacher's output at various time-steps and guidance scales. The training process involved 150K Unsplash images at 1024px resolution, running for 6,000 steps with a global batch size of 32. The training took approximately 12 hours with limited computational resources.
- Student model initialized with teacher parameters except for w-embedding
- Guidance scale matching in range w ∈ [1, 4]
- Uses true CFG instead of distilled CFG
- Requires custom inference script due to diffusers pipeline incompatibility
Core Capabilities
- True classifier-free guidance implementation
- High-resolution image processing (1024px)
- Flexible guidance scale control
- Compatible with custom inference implementations
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its approach to de-distilling guidance, effectively reversing the traditional distillation process to achieve true classifier-free guidance. This makes it particularly interesting for researchers and developers working on diffusion models.
Q: What are the recommended use cases?
The model is best suited for research purposes and applications requiring true classifier-free guidance. It's particularly useful for those interested in studying the differences between distilled and true CFG implementations in diffusion models.