ComfyUI: Difference between revisions

From Hegemon Wiki
Jump to navigation Jump to search
(Created page with "== Benchmark == {| class="wikitable" |+ ! ! |- |Using xformers |02:51, 42.77s/it |- | | |- | | |}")
 
 
(73 intermediate revisions by 4 users not shown)
Line 1: Line 1:
== TODO ==
Upscale Fix?

Tiled Diffusion?

[https://www.youtube.com/watch?v=dG9HfgngsqM Detail Daemon]

IPAdapter

Outpainting - https://www.youtube.com/watch?v=j20P4hAZS1Q https://www.youtube.com/watch?v=qLZJ7iSq9tY

== Misc ==
[https://www.youtube.com/watch?v=E_h78GRh3Rw Regional prompting for AI Image Generation (Flux and ComfyUI)] - Both DetailD (including the custom sampler stuff) and region based prompting

[https://www.youtube.com/watch?v=jgimrovxT3E ComfyUI Tutorial, SEGS Picker: How to prompt for each individual face]

[https://www.youtube.com/watch?v=xHvtwqtOhgE EASY Outpainting in ComfyUI: 3 Simple Ways with Auto-Tagging (Joytag) | Creative Workflow Tutorial]

https://www.youtube.com/@goshniiAI/videos - Upscaling

[https://www.youtube.com/watch?v=pYqMQwBbECk Generate Million-Megapixel Images with Z Image Turbo!]

[https://www.youtube.com/watch?v=Kmve1_jiDpQ How to Train a Z-Image-Turbo LoRA with AI Toolkit]

[https://www.youtube.com/watch?v=k0rzfhZM58s Stop Making Blurry AI Art! The "Halftone" Fix for Wan & Qwen (ComfyUI Workflow)]

[https://www.youtube.com/watch?v=XVuN4igXdoQ Insane ComfyUI Upgrade: 165 Custom Nodes That Actually Work]

https://openart.ai/

[https://www.youtube.com/watch?v=dASMuSCCNSU Different Region = Different LoRA! New ComfyUI Nodes for Area Composition]

[https://www.youtube.com/watch?v=Wv_O_Us4NdY Comfyui 101: How to Create X/Y Plots for SDXL & Flux] [https://www.youtube.com/watch?v=z_SJOr25lVo How to Use Flux Unet ( & GGUF) Models With TinyTerraNodes XY Plot]

[https://www.youtube.com/playlist?list=PL-pohOSaL8P9kLZP8tQ1K1QWdZEgwiBM0 ComfyUI Tutorial Series]

[https://civitaiarchive.com/ https://civitaiarchive.com] [https://civitasbay.org/?sort_by=downloads https://civitasbay.org/]

https://tensor.art/

https://www.liblib.art/

[https://huggingface.co/wsj1995/Checkpoint/tree/main Bunch of unsorted loras]

[https://www.youtube.com/watch?v=E5ngsV4OPBM Qwen Image Edit]

[https://github.com/jitcoder/lora-info lora-info]

[https://github.com/willmiao/ComfyUI-Lora-Manager ComfyUI-Lora-Manager]

[https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main zer0int/CLIP-GmP-ViT-L-14/tree/main] - what's the best version of Vit-L clip encoder? smooth and detail both r good

Updateing

git pull

== Installing ==
uv pip install --upgrade torch torchvision torchaudio --index-url <nowiki>https://download.pytorch.org/whl/cu130</nowiki>

uv pip install sageattention==2.2.0 --no-build-isolation


pip install -r requirements.txt

== Prompts ==
[X : Y : N] - [https://civitai.com/articles/4808/making-images-great-again-remastered Use X for N steps then Y]. Not sure if this works in comfyui.

[https://ltx.io/model/model-blog/prompting-guide-for-ltx-2 Prompting Guide for LTX-2]

[https://docs.bfl.ai/guides/prompting_guide_flux2_klein <nowiki>Prompting Guide - FLUX.2 [klein]</nowiki>]

https://nowprompts.com/

https://docs.bfl.ai/guides/prompting_guide_kontext_i2i

https://danbooru.donmai.us/wiki_pages/tag_groups

https://danbooru.donmai.us/related_tag

https://danbooru.donmai.us/posts?tags=list_of_style_parodies&z=2

https://betterwaifu.com/blog/ai-pose-prompts

=== Artists ===
Deak Ferrand - Dune, Sandman

[https://civitai.com/images/6194935 by Diane Pierce and Alena Aenami in the style of Dan Mumford,]

Asaf hanuka

rodolphe wytsman impressionism

by Michal Karcz, by Arthur Sarnoff, by Marc Simonetti,

Greg Rutkowski

https://aiartes.com/artists

[https://www.youtube.com/watch?v=0fRk6RH-zTw Peter Elson]

Syd Mead

==== Cyberpunk ====
Syd Mead - Bladerunner concept art

Josan Gonzalez

Sparth (Nicolas Bouvier)

Beeple (Mike Winkelmann) - Weird colours

Simon Stålenhag - Giant robots guy

==== Fantasy ====
Greg Rutkowski

== Benchmark ==
== Benchmark ==


=== Video Upscale ===


[https://www.reddit.com/r/StableDiffusion/comments/1opb9wk/seedvr2_comfyui_4x_upscale_poor_performance_on_a/#:~:text=The%20nightly%20version%20of%20SeedVR2,faster%20and%20more%20memory%20efficient. FlashVSR]
{| class="wikitable"
{| class="wikitable"
|+
|+
!
!
!
!
!
!
!
|-
|-
|2026-01-30
|Using xformers
|336.80s
|02:51, 42.77s/it
|GAN upscaler template in ColmfyUI
|for a 3 second video using the
|
|-
|-
|2026-01-30
|19.04s
|GAN upscaler template in ColmfyUI
|for a second pass (Cached or load time?)
|
|-
|2026-01-30
|343.10s
|'''seedvr2_ema_3b_fp16'''
|
|
|
|
|-
|-
|2026-01-30
|351.37
|'''seedvr2_ema_3b_fp8'''
|
|
|-
|2026-01-30
|354.54
|seedvr2_ema_3b-Q4_K_M
|
|
|-
|2026-01-30
|442.38
|seedvr2_ema_3b_fp8
|With torch compile and tile size of '''768''' (First run)
|
|-
|2026-01-30
|305.02
|seedvr2_ema_3b_fp8
|With torch compile and tile size of 768 ('''Second''' run)
|
|-
|2026-01-30
|408.12
|seedvr2_ema_3b_fp8
|With torch compile and tile size of 768, '''swap blocks 6''' ('''First''' run)
|
|-
|2026-01-30
|341.21
|seedvr2_ema_3b_fp8
|With torch compile and tile size of 768, '''swap blocks 6,''' batch size 21
|
|-
|2026-01-30
|375.13
|seedvr2_ema_3b_fp8
|With torch compile and encoding tile size of '''1024''', '''swap blocks 6,''' batch size 21
|
|-
|2026-01-30
|323.73
|seedvr2_ema_3b_fp8
|With torch compile and encoding tile size of '''1024''', '''swap blocks 9,''' batch size 21, sageattention2
|
|-
|2026-01-30
|405.08
|seedvr2_ema_7b_Q4_K_M
|Default settings
|
|-
|
|
|
|
|
|
|
|}
|}

== Training LORAs ==
All the training programs just seem to be wrappers around hugging face scripts.
{| class="wikitable"
|+
!
!Last Updated
!⭐
!Z-Image
!Flux.2 Klein
!Qwen
|-
|[https://github.com/bmaltais/kohya_ss Kohya SS]
|7 Months Ago
|11.9k
|
|
|
|-
|[https://github.com/Nerogar/OneTrainer OneTrainer]
|2 days ago.
|2.7k
|✔️ (unknown base/turbo)
|❌?
|✔️
|-
|[https://github.com/ostris/ai-toolkit Ai Toolkit]
|2 days ago.
|9.2k
|
|
|
|-
|[https://github.com/kohya-ss/musubi-tuner Musubi]
|Yesterday.
|1.7k
|✔️ (base)
|✔️
|✔️ (including layered)
|}

== Seed VR2 ==
[https://www.youtube.com/watch?v=LT6mw1iAqkQ SeedVR2 BEST SETTINGS for Every GPU! (6GB–32GB+) Nobody Told You This…]

[https://www.youtube.com/watch?v=MBtWYXq_r60 🤿 SeedVR2 v2.5 Video Upscaling: Official Guide from the ComfyUI Integration Team | AInVFX Nov 7] - Talks about out of memory at different phases.

== LTX-2 ==
[https://www.youtube.com/watch?v=AgF5-98-f6E LTX 2 Q6 16GB on 8GB VRAM? 🤯 Free ComfyUI Workflow 2026]

[https://www.reddit.com/r/StableDiffusion/comments/1qe9ub2/ltx2_generate_a_30s_video_in_310seconds/ LTX-2 generate a 30s video in 310seconds] - Has distilled GGUFs

[https://www.youtube.com/watch?v=-js3Lnq3Ip4 LTX-2 Simplified Workflow 🔥 Distilled Checkpoints or Separated VAE & Transformer?]

[https://www.youtube.com/watch?v=5-cYtn4dAH0 LTX-2 Fix: Yes, You Can Actually Use It Now (16-24gb VRAM)] - Didn't work

[https://www.reddit.com/r/StableDiffusion/comments/1q9utj2/this_fixed_my_oom_issues_with_ltx2/ This fixed my OOM issues with LTX-2] - Didn't work.

== Ace Step 1.5 ==
[https://civitai.com/models/1558969?modelVersionId=1832664 Civit AI Workflow] - "Cover - Style transfer that preserves the semantic structure (rhythm, melody) while generating new audio with different characteristics"

Latest revision as of 13:43, 14 February 2026

TODO[edit | edit source]

Upscale Fix?

Tiled Diffusion?

Detail Daemon

IPAdapter

Outpainting - https://www.youtube.com/watch?v=j20P4hAZS1Q https://www.youtube.com/watch?v=qLZJ7iSq9tY

Misc[edit | edit source]

Regional prompting for AI Image Generation (Flux and ComfyUI) - Both DetailD (including the custom sampler stuff) and region based prompting

ComfyUI Tutorial, SEGS Picker: How to prompt for each individual face

EASY Outpainting in ComfyUI: 3 Simple Ways with Auto-Tagging (Joytag) | Creative Workflow Tutorial

https://www.youtube.com/@goshniiAI/videos - Upscaling

Generate Million-Megapixel Images with Z Image Turbo!

How to Train a Z-Image-Turbo LoRA with AI Toolkit

Stop Making Blurry AI Art! The "Halftone" Fix for Wan & Qwen (ComfyUI Workflow)

Insane ComfyUI Upgrade: 165 Custom Nodes That Actually Work

https://openart.ai/

Different Region = Different LoRA! New ComfyUI Nodes for Area Composition

Comfyui 101: How to Create X/Y Plots for SDXL & Flux How to Use Flux Unet ( & GGUF) Models With TinyTerraNodes XY Plot

ComfyUI Tutorial Series

https://civitaiarchive.com https://civitasbay.org/

https://tensor.art/

https://www.liblib.art/

Bunch of unsorted loras

Qwen Image Edit

lora-info

ComfyUI-Lora-Manager

zer0int/CLIP-GmP-ViT-L-14/tree/main - what's the best version of Vit-L clip encoder? smooth and detail both r good

Updateing

git pull

Installing[edit | edit source]

uv pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

uv pip install sageattention==2.2.0 --no-build-isolation


pip install -r requirements.txt

Prompts[edit | edit source]

[X : Y : N] - Use X for N steps then Y. Not sure if this works in comfyui.

Prompting Guide for LTX-2

Prompting Guide - FLUX.2 [klein]

https://nowprompts.com/

https://docs.bfl.ai/guides/prompting_guide_kontext_i2i

https://danbooru.donmai.us/wiki_pages/tag_groups

https://danbooru.donmai.us/related_tag

https://danbooru.donmai.us/posts?tags=list_of_style_parodies&z=2

https://betterwaifu.com/blog/ai-pose-prompts

Artists[edit | edit source]

Deak Ferrand - Dune, Sandman

by Diane Pierce and Alena Aenami in the style of Dan Mumford,

Asaf hanuka

rodolphe wytsman impressionism

by Michal Karcz, by Arthur Sarnoff, by Marc Simonetti,

Greg Rutkowski

https://aiartes.com/artists

Peter Elson

Syd Mead

Cyberpunk[edit | edit source]

Syd Mead - Bladerunner concept art

Josan Gonzalez

Sparth (Nicolas Bouvier)

Beeple (Mike Winkelmann) - Weird colours

Simon Stålenhag - Giant robots guy

Fantasy[edit | edit source]

Greg Rutkowski

Benchmark[edit | edit source]

Video Upscale[edit | edit source]

FlashVSR

2026-01-30 336.80s GAN upscaler template in ColmfyUI for a 3 second video using the
2026-01-30 19.04s GAN upscaler template in ColmfyUI for a second pass (Cached or load time?)
2026-01-30 343.10s seedvr2_ema_3b_fp16
2026-01-30 351.37 seedvr2_ema_3b_fp8
2026-01-30 354.54 seedvr2_ema_3b-Q4_K_M
2026-01-30 442.38 seedvr2_ema_3b_fp8 With torch compile and tile size of 768 (First run)
2026-01-30 305.02 seedvr2_ema_3b_fp8 With torch compile and tile size of 768 (Second run)
2026-01-30 408.12 seedvr2_ema_3b_fp8 With torch compile and tile size of 768, swap blocks 6 (First run)
2026-01-30 341.21 seedvr2_ema_3b_fp8 With torch compile and tile size of 768, swap blocks 6, batch size 21
2026-01-30 375.13 seedvr2_ema_3b_fp8 With torch compile and encoding tile size of 1024, swap blocks 6, batch size 21
2026-01-30 323.73 seedvr2_ema_3b_fp8 With torch compile and encoding tile size of 1024, swap blocks 9, batch size 21, sageattention2
2026-01-30 405.08 seedvr2_ema_7b_Q4_K_M Default settings

Training LORAs[edit | edit source]

All the training programs just seem to be wrappers around hugging face scripts.

Last Updated Z-Image Flux.2 Klein Qwen
Kohya SS 7 Months Ago 11.9k
OneTrainer 2 days ago. 2.7k ✔️ (unknown base/turbo) ❌? ✔️
Ai Toolkit 2 days ago. 9.2k
Musubi Yesterday. 1.7k ✔️ (base) ✔️ ✔️ (including layered)

Seed VR2[edit | edit source]

SeedVR2 BEST SETTINGS for Every GPU! (6GB–32GB+) Nobody Told You This…

🤿 SeedVR2 v2.5 Video Upscaling: Official Guide from the ComfyUI Integration Team | AInVFX Nov 7 - Talks about out of memory at different phases.

LTX-2[edit | edit source]

LTX 2 Q6 16GB on 8GB VRAM? 🤯 Free ComfyUI Workflow 2026

LTX-2 generate a 30s video in 310seconds - Has distilled GGUFs

LTX-2 Simplified Workflow 🔥 Distilled Checkpoints or Separated VAE & Transformer?

LTX-2 Fix: Yes, You Can Actually Use It Now (16-24gb VRAM) - Didn't work

This fixed my OOM issues with LTX-2 - Didn't work.

Ace Step 1.5[edit | edit source]

Civit AI Workflow - "Cover - Style transfer that preserves the semantic structure (rhythm, melody) while generating new audio with different characteristics"