Skip to content

simple diffusers based implementation of Hunyuan-DiT, in Forge webUI for Stable Diffusion. Works with 8GB VRAM.

Notifications You must be signed in to change notification settings

DenOfEquity/Hunyuan-DiT-for-webUI

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hunyuan-DiT for webui

Forge tested, probably A1111 too

I don't think there is anything Forge specific here.

works for me TM on 8GB VRAM, 16GB RAM (GTX1070)


Install

Go to the Extensions tab, then Install from URL, use the URL for this repository.

Important

needs diffusers 0.30.0 or newer

Easiest way to ensure suitably up-to-date diffusers is installed is to edit requirements_versions.txt in the webUI folder.

diffusers>=0.30.0

You might also need to update:

tokenizers>=0.19
transformers>=4.40
huggingface-hub>=0.23.4

Forge2 already has newer versions for all.


screenshot

not current UI, but similar


downloads models on first run - ~13.4GB (optionally +5.6GB for variants)


Change log

??/08/2024

  • adjustments for Gradio4

07/08/2024

  • diffusers 0.30.0 finally released, so now we have:
    • v1.2 of HunyuanDiT (distilled variant as default)
    • controlnet - will fit into 8GB VRAM (actually ~7GB), but you'd best not be doing much else at the same time. If inference becomes very slow or fails, you've run out of VRAM. Control images must be preprocessed, you can use controlnet in the main txt2img tab for this.
    • whatever else I added/fixed/broke in the meantime

27/07/2024

  • added drawing of masks for image to image. Load/copy the source image into the mask, to use as a template.
  • added option to keep models loaded

24/07/2024

  • added SuperPrompt button (ꌗ) to rewrite simple prompts with more detail. This overwrites the prompt. Read about SuperPrompt here. Credit to BrianFitzgerald for the model. (all my alternate model extensions are updated to use this; the model is loaded to a shared location so there's no wasted memory due to duplicates.)

10/07/2024

  • improved yesterday's effort. More compatibility, multi-line, etc.

09/07/2024

  • some code cleanups
  • added prompt parsing to automatically fill in details like seed, steps, etc.

05/07/2024

  • added guidance cut off control: switch to CFG 1.0 after some steps for faster generation. The earlier the switch, the more effect on quality. A late switch might not hurt quality at all.

04/07/2024

  • added toggle button to center the latents during inference. This is something that can be very effective with sd1.5 and sdXL (but not sd3). Tends to change the results significantly with HYdit, but doesn't break them, so it's included for experimentation.

03/07/2024

  • added quickset for image resolution, included sizes are the recommended ones (and would be enforced if resolution binning were enabled, which it isn't).
  • tweaked Florence-2: model now runs on GPU so is faster.

02/07/2024

  • cleaner handling for batch after understanding the code a bit better - why repeat over the second dimension, then take use a view to get the right shape, when it is easier to just repeat over the first dimension? Other diffusers pipelines do it too.
  • fixed image to image: changed method of added noise when I added masking, which was an error.
  • removed Euler and Euler A samplers as they seem very bad.

01/07/2024

  • fix for batch, tested with diffusers versions 0.29.1 and 0.29.2. On my low-end hardware, this takes n times longer to generate, up to the point where it starts hitting shared memory, then it gets really slow.
  • model selection now via dropdown, in preparation for v1.2 (not much information yet, needs a code update, so still have v1.1 as default)
  • (controlnet: waiting for diffusers update)

29/06/2024

  • added lora support, Tencent original form (with strength control) and pre-converted diffusers form (ignores strength control). Of course, there are already alternate form loras available which I can't load. Some sort of conversion script may be possible. Generation will end early, with a console message, for those loras.
  • mask for image to image

25/06/2024

  • added option to caption using Florence-2, in image to image section. 'P' button toggles overwriting prompt, results always written to console.
  • minor code improvements

17/06/2024

  • minor addition to save noise colour settings to infotext

16/06/2024

  • added v1.1. Enabled by default, but optional by using the obvious button. I think only the transformer has changed, so 5.64GB extra download (same for Distilled, if used). From brief tests, it does seem to be a step-up.
  • option to not use T5 text encoder
  • settings to colourize the initial noise. This offers some extra control over the output and is near-enough free. Leave strength at 0.0 to bypass it.
  • experimental double prompting - subprompts for each text encoder. Split prompts with '|', first subprompt for CLIP, second for T5. If not used: same prompt sent to both, same as previous behaviour.

12/06/2024

  • code cleanup, handles the text encoders manually, better for VRAM usage. In good conditions, no speed up; but bad conditions are harder to hit.
  • moved styles to unique file

08/06/2024

  • added support for the distilled version, which is better when using fewer steps. I download only the distilled transformer, so the cost is ~5.6GB rather than another 13GB. Toggle the D icon, top-right of left column: lit up means using the distilled version. Downloaded on demand, cached locally. Should this be new default?

07/06/2024

  • !! don't apply i2i denoise strength when not doing i2i, late night me forgot to copy that over from the PixArt implementation
  • enabled guidance rescale for testing. It's good, very similar method used in my cfgFade extension.

05/06/2024

  • reduced VRAM, no longer flirting with shared memory
  • caching of prompt embeds to avoid text encoder processing if prompt and negative not changed
  • img2img, same method as used with PixArt

04/06/2024

Initial release, dips into shared memory too easily. K icon (top-right of left column) toggles use of Karras sigmas for the samplers. Seemed useful with PixArt + Cascade, so why not here?


image2image progression with a nice denoise


example for cfg, steps, rescale cfg

Of course, always same prompt, seed, sampler. Non-distilled version.

From left to right:

  • cfg: 2, 4, 8, 8
  • steps: 20, 20, 20, 40 Top row: 0 rescale; bottom row: 0.75 rescale


Generating with 8GB VRAM is possible. Using CFG 1 saves some VRAM and is considerably faster, but still slower than equivalent resolutions with sdXL or PixArt. Using small resolutions (768x768) seems to give very poor/broken results. Resolution binning is NOT enabled (width/height would be automatically adjusted to 'supported' values) as this seems to cause issues along borders.


example

prompt: photograph of a kintsugi bowl of steaming dumplings on a rustic wooden table

About

simple diffusers based implementation of Hunyuan-DiT, in Forge webUI for Stable Diffusion. Works with 8GB VRAM.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages