Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

1.10.0-RC#16158

Jul 6, 2024· 12 comments· 12 replies
Discussion options

How to switch to different versions of WebUI

Features:

  • A lot of performance improvements (see below in Performance section)
  • Stable Diffusion 3 support (#16030)
    • Recommended Euler sampler; DDIM and other timestamp samplers currently not supported
    • T5 text model is disabled by default, enable it in settings
  • New schedulers:
  • New sampler: DDIM CFG++ (#16035)

Minor:

  • Option to skip CFG on early steps (#15607)
  • Add --models-dir option (#15742)
  • Allow mobile users to open context menu by using two fingers press (#15682)
  • Infotext: add Lora name as TI hashes for bundled Textual Inversion (#15679)
  • Check model's hash after downloading it to prevent corruped downloads (#15602)
  • More extension tag filtering options (#15627)
  • When saving AVIF, use JPEG's quality setting (#15610)
  • Add filename pattern:[basename] (#15978)
  • Add option to enable clip skip for clip L on SDXL (#15992)
  • Option to prevent screen sleep during generation (#16001)
  • ToggleLivePriview button in image viewer (#16065)

Extensions and API:

  • Add process_before_every_sampling hook (#15984)
  • Return HTTP 400 instead of 404 on invalid sampler error (#16140)

Performance:

  • [Performance 1/6] use_checkpoint = False (#15803)
  • [Performance 2/6] Replace einops.rearrange with torch native ops (#15804)
  • [Performance 4/6] Precompute is_sdxl_inpaint flag (#15806)
  • [Performance 5/6] Prevent unnecessary extra networks bias backup (#15816)
  • [Performance 6/6] Add --precision half option to avoid casting during inference (#15820)
  • [Performance] LDM optimization patches (#15824)
  • [Performance] Keep sigmas on CPU (#15823)
  • Check for nans in unet only once, after all steps have been completed
  • Added pption to run torch profiler for image generation

Bug Fixes:

  • Fix for grids without comprehensive infotexts (#15958)
  • feat: lora partial update precede full update (#15943)
  • Fix bug where file extension had an extra '.' under some circumstances (#15893)
  • Fix corrupt model initial load loop (#15600)
  • Allow old sampler names in API (#15656)
  • more old sampler scheduler compatibility (#15681)
  • Fix Hypertile xyz (#15831)
  • XYZ CSV skipinitialspace (#15832)
  • fix soft inpainting on mps and xpu, torch_utils.float64 (#15815)
  • fix extention update when not on main branch (#15797)
  • update pickle safe filenames
  • use relative path for webui-assets css (#15757)
  • When creating a virtual environment, upgrade pip in webui.bat/webui.sh (#15750)
  • Fix AttributeError (#15738)
  • use script_path for webui root in launch_utils (#15705)
  • fix extra batch mode P Transparency (#15664)
  • use gradio theme colors in css (#15680)
  • Fix dragging text within prompt input (#15657)
  • Add correct mimetype for .mjs files (#15654)
  • QOL Items - handle metadata issues more cleanly for SD models, Loras and embeddings (#15632)
  • replace wsl-open with wslpath and explorer.exe (#15968)
  • Fix SDXL Inpaint (#15976)
  • multi size grid (#15988)
  • fix Replace preview (#16118)
  • Possible fix of wrong scale in weight decomposition (#16151)
  • Ensure use of python from venv on Mac and Linux (#16116)
  • Prioritize python3.10 over python3 if both are available on Linux and Mac (with fallback) (#16092)
  • stoping generation extras (#16085)
  • Fix SD2 loading (#16078,#16079)
  • fix infotext Lora hashes for hires fix different lora (#16062)
  • Fix sampler scheduler autocorrection warning (#16054)

Other:

This discussion was created from the release1.10.0-RC.
You must be logged in to vote

Replies: 12 comments 12 replies

Comment options

Hires.fix not working with sd3 model (--medvram)

File "C:\sd.webui\webui\modules\models\sd3\sd3_impls.py", line 75, in apply_model
return self.model_sampling.calculate_denoised(sigma, model_output, x)
File "C:\sd.webui\webui\modules\models\sd3\sd3_impls.py", line 43, in calculate_denoised
return model_input - model_output * sigma
RuntimeError: The size of tensor a (123) must match the size of tensor b (122) at non-singleton dimension 3

SD3 lora from civitai not working

  File "C:\sd.webui\webui\modules\models\sd3\mmdit.py", line 441, in forward    return block_mixing(*args, context_block=self.context_block, x_block=self.x_block, **kwargs)  File "C:\sd.webui\webui\modules\models\sd3\mmdit.py", line 410, in block_mixing    context_qkv, context_intermediates = context_block.pre_attention(context, c)  File "C:\sd.webui\webui\modules\models\sd3\mmdit.py", line 384, in pre_attention    qkv = self.attn.pre_attention(modulate(self.norm1(x), shift_msa, scale_msa))  File "C:\sd.webui\webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1518, in _wrapped_call_impl    return self._call_impl(*args, **kwargs)  File "C:\sd.webui\webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1527, in _call_impl    return forward_call(*args, **kwargs)  File "C:\sd.webui\webui\extensions-builtin\Lora\networks.py", line 572, in network_LayerNorm_forward    network_apply_weights(self)  File "C:\sd.webui\webui\extensions-builtin\Lora\networks.py", line 397, in network_apply_weights    weights_backup = self.weight.to(devices.cpu, copy=True)AttributeError: 'NoneType' object has no attribute 'to'
You must be logged in to vote
0 replies
Comment options

any benchmarking how much % faster it would be, comparing to 1.9?

You must be logged in to vote
1 reply
@light-and-ray
Comment options

Different systems show different performance boost

Comment options

/home/----/stable-diffusion-webui/modules/sd_samplers_common.py:68: RuntimeWarning: invalid value encountered in cast
x_sample = x_sample.astype(np.uint8)
This error at hires pictures. I use WSL, AMD gpu ROCm.
00476-4089085403
sub-quadratic optimize used. SDP breaking at heap memory error at 600*900 2x hires.

You must be logged in to vote
0 replies
Comment options

Thank you!

You must be logged in to vote
0 replies
Comment options

why not implemented supportCosXL ?

You must be logged in to vote
0 replies
Comment options

With this, is performance for SDXL on par with sd-forge? In their deviation announcement, they mentioned a1111 should have implemented all their performance improvements for v1.10

You must be logged in to vote
1 reply
@HMRMike
Comment options

It's definitely faster than before. Also the newskip_early_cond will boost speed in the beginning, depending on your setting, and has its own quality benefits. I still think forge was faster but this isn't a significant difference now. The memory management though, it's where Forge really shines still.

Comment options

How do I enable Option to skip CFG on early steps (#15607)
and if i upgrade from current install do i need to enable anything in settings to get the other performance upgrades?

You must be logged in to vote
4 replies
@Ineman
Comment options

setting-interface-quick setting
add skip_early_cond
restart ui

@HMRMike
Comment options

For the performance improvements, I think we only need to set a new commandline arg as explained here, if your GPU is compatible:
#15821
The other upgrades just work on their own. I had an issue where setting '--precision half' actually dropped performance to 2it/s (SDXL, euler, 1024) on 3090. Gave up and just installed from scratch (save your styles file!). Seems to work fine now.

@NulliferBones
Comment options

Alright thank you.

Something i noticed and don't know if its intended behavior, but if you try to set skip cfg and load a model that has an inpainting flag, it will not generate anything at all in any tab.

Also out of curiosity if anyone knows. I have 64GB ram, and a GTX 1060 6GB. Xformers I get 1.2 S/it with realistic vision, DPM++ SDE with Align your steps. 512x512
768x768 is roughly 3-3.5 S/it
Hires fix steps on 512 upscale 2 are roughly 7-8 S/it

Do you think I would see any worth while improvements by moving over to linux?

Also I OOM on final step of hires fix 2x at 768 on this new version

@NulliferBones
Comment options

short story, yes i get some stability and performance improvement

Comment options

When I try to use SDXL models/loras and 30+ steps it generates a blurry pixelated image. The first steps seem fine with the proper image but then suddely it becomes a pixelated mess. It works fine with 25 steps.

Also, using much more RAM now. Before it was 42GB now it is 82GB.

Example:
Capturar

You must be logged in to vote
3 replies
@Takezo1000
Comment options

After some testing the source of the problem seems to be the Euler a sampler with the automatic and uniform scheduler. It works normally with other schedulers like Karras

@Takezo1000
Comment options

After some testing with the arguments I removed--no-half --precision full and the RAM usage went down from 72GB to 40-42GB.

Before: set COMMANDLINE_ARGS=--use-directml --theme dark --medvram-sdxl --medvram --no-half --precision full --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --autolaunch
set SAFETENSORS_FAST_GPU=1

After: set COMMANDLINE_ARGS=--use-directml --theme dark --medvram-sdxl --medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --autolaunch
set SAFETENSORS_FAST_GPU=1

@light-and-ray
Comment options

You use different project COMMANDLINE_ARGS=--use-directml

Comment options

Its a bit confusing that SDXL LoRAs will be listed when SD3 is loaded and when I add a specific SD3 LoRA I get the following error "AttributeError: 'NoneType' object has no attribute 'to'. I presume there is no LoRA support in this RC?

You must be logged in to vote
0 replies

This comment has been hidden.

@w-e-w

This comment has been hidden.

Comment options

I'm not sure what the advantage of this version numbering is, but in general terms, if name the next version after "1.9.4" as "2.0.0" instead of "1.10.0", it will be easier to understand because it will fit into three digits instead of four.

You must be logged in to vote
2 replies
@elistys
Comment options

Would it be easier for the general public to understand if it was named "Stable Diffusion 1.30" rather than "Stable Diffusion 3"?

@Soulreaver90
Comment options

You are just trolling now aren't you? SD 1.3 existed, it came out before 1.4 and the commonly used 1.5. Plus SD3 is the official name by Stability AI.

Comment options

If we compare command line parameters with those from forge--precision half is the same thing as--all-in-fp16. But is there any counterpart of--attention-pytorch and--always-high-vram?

You must be logged in to vote
0 replies
Sign up for freeto join this conversation on GitHub. Already have an account?Sign in to comment
Labels
None yet
18 participants
@AUTOMATIC1111@HMRMike@tritant@leporel@EaglePPP@Cruxial0@viking1304@light-and-ray@dkspwndj@w-e-w@Ineman@Soulreaver90@NulliferBones@steve84wien@elistys@guythis31@Takezo1000@DotPoker2
[8]ページ先頭
©2009-2025 Movatter.jp