List of "Hidden" Tricks

List of "Hidden" Tricks

The below things are already inside the software, and users do not need to do anything about these.

  1. GPT2-based prompt expansion as a dynamic style "Fooocus V2". (similar to Midjourney's hidden pre-processsing and "raw" mode, or the LeonardoAI's Prompt Magic).

  2. Native refiner swap inside one single k-sampler. The advantage is that the refiner model can now reuse the base model's momentum (or ODE's history parameters) collected from k-sampling to achieve more coherent sampling. In Automatic1111's high-res fix and ComfyUI's node system, the base model and refiner use two independent k-samplers, which means the momentum is largely wasted, and the sampling continuity is broken. Fooocus uses its own advanced k-diffusion sampling that ensures seamless, native, and continuous swap in a refiner setup. (Update Aug 13: Actually, I discussed this with Automatic1111 several days ago, and it seems that the โ€œnative refiner swap inside one single k-samplerโ€ is merged into the dev branch of webui. Great!)

  3. Negative ADM guidance. Because the highest resolution level of XL Base does not have cross attentions, the positive and negative signals for XL's highest resolution level cannot receive enough contrasts during the CFG sampling, causing the results to look a bit plastic or overly smooth in certain cases. Fortunately, since the XL's highest resolution level is still conditioned on image aspect ratios (ADM), we can modify the adm on the positive/negative side to compensate for the lack of CFG contrast in the highest resolution level. (Update Aug 16, the IOS App Drawing Things will support Negative ADM Guidance. Great!)

  4. We implemented a carefully tuned variation of Section 5.1 of "Improving Sample Quality of Diffusion Models Using Self-Attention Guidance". The weight is set to very low, but this is Fooocus's final guarantee to make sure that the XL will never yield an overly smooth or plastic appearance (examples here). This can almost eliminate all cases for which XL still occasionally produces overly smooth results, even with negative ADM guidance. (Update 2023 Aug 18, the Gaussian kernel of SAG is changed to an anisotropic kernel for better structure preservation and fewer artifacts.)

  5. We modified the style templates a bit and added the "cinematic-default".

  6. We tested the "sd_xl_offset_example-lora_1.0.safetensors" and it seems that when the lora weight is below 0.5, the results are always better than XL without lora.

  7. The parameters of samplers are carefully tuned.

  8. Because XL uses positional encoding for generation resolution, images generated by several fixed resolutions look a bit better than those from arbitrary resolutions (because the positional encoding is not very good at handling int numbers that are unseen during training). This suggests that the resolutions in UI may be hard coded for best results.

  9. Separated prompts for two different text encoders seem unnecessary. Separated prompts for the base model and refiner may work, but the effects are random, and we refrain from implementing this.

  10. The DPM family seems well-suited for XL since XL sometimes generates overly smooth texture, but the DPM family sometimes generates overly dense detail in texture. Their joint effect looks neutral and appealing to human perception.

  11. A carefully designed system for balancing multiple styles as well as prompt expansion.

  12. Using automatic1111's method to normalize prompt emphasizing. This significantly improves results when users directly copy prompts from civitai.

  13. The joint swap system of the refiner now also supports img2img and upscale in a seamless way.

  14. CFG Scale and TSNR correction (tuned for SDXL) when CFG is bigger than 10.

Last updated