🏆🏆🏆 MY BOUNTY IS LIVE 🏆🏆🏆 Looking for BEST images for Cinematic Realism!
https://civarchive.com/bounties/10855
This workflow UNLOCKS the full potential of Z-Image. It generates pixel-level fine realistic details with very fast gen speed, for both foreground and backgrounds!
I spent months researching and experimenting Qwen hyper realism and now weeks of chasing after Z Image realism, and I'm now sharing my full workflow with this release, you might find it useful to use.
Thanks for the shoutout by Z Image dev team TongyiLab!! https://x.com/Ali_TongyiLab/status/2000762326136578121
v2.2 Key features: (see compare pics in v2.2 post, not going to make v2.3 until I find another significant image quality boost technique)
NEW Detail vs Cinematic master controls. Now users have full control over this!
NEW Quick reuse prompt section, one click regen image with new seeds
NEW Figured out Z Image fp32 noise issue, attached fix suggestions
Achieve realism across broad range of fictional/nonfictional prompts
Get high details up close AND far away in background
Maintains realistic skin/texture through main image and through upscale
Very fast initial generation, upscale optional.
Uses my Loras to achieve higher detail density & improved composition
Built in high quality+fast dual workflow/stand-alone upscaler
High quality Upscaled Inpainter
Realistic PBR like having nice specularity and sharpness of edges and good metalness
Easy to use workflow that won't require per-image tweaking
Photographer style Neutral image output with max color range
Clean 3D structure such as building lines and objects
❤And if you want to incorporate some parts into your workflow, I would love a mention too! 🏆MY CINE REALISM BOUNTY IS LIVE: https://civarchive.com/bounties/10855
Note: This workflow works much better with my Cinematic and CyberHD Enhancer loras since they increase the resolution of the details and textures.
FLOWMATCH NOISE FIX: ComfyUI beyond 0.3.77 changed the default UNET torch weight dType, for many users it used to default to fp32. FlowMatch is sensitive to this, and works best with fp32. To set the toggle, edit your .bat file for launching ComfyUI, and add to the end: --fp32-unet
Comparison article here: https://civarchive.com/articles/24424
IF YOU ARE NOT GETTING MY LEVEL OF DETAIL/NOISY IMAGE:
It probably means your Flowmatch isn't working. You MAY get this error log: The config attributes {'stochastic_sampling': False} were passed to FlowMatchEulerDiscreteScheduler, but are not expected and will be ignored. Please verify your scheduler_config.json configuration file.
(I had to downgrade to ComfyUI 0.3.77 to work)
(SECURITY ALERT: DEC 29: FIXED in my Workflow v2.01 REMOVED: RMBG nodes from AILAB. Security Vuln in nodes.
Usage Guide:
Just toss in a prompt and it'll just work!
Let me know if you have any questions/guidance requested in comments.
Description
v2.01 (SECURITY ALERT: DEC 29: v2.01 REMOVED: RMBG nodes from AILAB. Security Vuln in nodes. (Replaced one image crop node in Inpainter section with KJNode instead)
v2.0 REVAMPED EVERYTHING, HAVE FUN! (Check all the notes in each section!)
Revamped Upscaler, now fully retains texture and details.
Revamped Main Samplers, now more cinematic AND more realistic at the same time.
FAQ
Comments (33)
🚸🚸🚸v2.01 (SECURITY ALERT: DEC 29: v2.01 REMOVED: RMBG nodes from AILAB. Security Vuln in nodes. (Replaced one image crop node in Inpainter section with KJNode instead)
As of today, I scanned the code files and found the torch vulnerabilities in torch.load (Thanks to heads up from Scruffy and Kijai )
Is it same as https://github.com/1038lab/ComfyUI-RMBG/issues/151 or a different one?
@JzPz It's related. The initial discovery was from that, but the Commit around dec 14th only fixed 1 of the vulns but not the patched torch vuln, and there may be other nodes in that pack with the same problem. (I looked at the latest file today)
So is this workflow fixed and updated? Or should I remove this RMBG node and replaced with One Image crop node by Kijai KJNode?
@aneebtheartist126 In 2.01 I fixed the workflow already. RMBG is a series of very popular nodes actually =(
@luneva Oh ok let me try V2.01. Thanks Legend.
What is this? Thank you Mario! But the workflow you want to find is in another castle!
(check latest version number in top menu tab)
Which version is correct and updated ? There is no 2.0 only 2.01 . Which version fixed RMBG nodes from AILAB?
version 2.01 is the latest =) I needed to place any file to have this Menu Tab remain active, and I didn't want to keep updating the main tab here with new versions, I just want to use this old tab as a central place for image postings. The version select is up top on the menu!
Just checked, all versions before v2.01 contain RMBG nodes.
@DentoL yes correct, please update to v2.01!
OMG! OMG! OMG! You did it! You tamed the Shark detailler daemon with the workflow v2!! I shrieked like I was 12 years old at my first generation. Exactly what I had in my mind view. I had H.R. Giger effects with the workflow v1.9 and now it is all gone! Thank you, thank you.
thanks, it took quite some time! =) I actually need to release a v2.5, because I made a significant discovery to make the image even better now. But still doing final tuning to get it more consistent on skin texture.
FLOWMATCH NOISE FIX: ComfyUI beyond 0.3.77 changed the default UNET torch weight dType, for many users it used to default to fp32. FlowMatch is sensitive to this, and works best with fp32. To set the toggle, edit your .bat file for launching ComfyUI, and add to the end: --fp32-unet
Comparison article here: https://civitai.com/articles/24424
Thanks for the workflow and possible fix.
Have you tried using "--bf16-unet" instead of "--fp32-unet"? I wasn't able to run fp32 because I was running out of RAM and VRAM even when I have 16GB GPU and 32GB RAM. In my opinion it gets rid of the noise as well.
@Gh0sTt I was JUST about to try bf16 today, thanks for confirming and sharing that =)
@Gh0sTt so there are some quality differences, I'm going to compare the various modes and post a compare pic and gen times
@luneva yeah, I wasn't able to compare them. But for those of us who are hardware limited, it's still an improvement for this workflow
@Gh0sTt To set the toggle, edit your .bat file for launching ComfyUI, and add to the end: --fp32-unet
Comparison article here: https://civitai.com/articles/24424
@Gh0sTt In my next version, I will create a setting that will allow you to select softer noise (for those not using fp32)
@luneva I have a suggestion that might be even better: What about using a ModelComputeDtype node right after the ZIT loads? That keeps the ZIT at FP32 while allowing the CLIP and VAE to stay at their default weights. If it works as intended, it saves everyone from having to modify their .bat file.
@Gh0sTt ohh let me check that thanks =)
@Gh0sTt thank you it worked, I'll integrate it into my next workflow version, which will come with some cleanup nodes and also improved image quality tuning.
@Gh0sTt in the new workflow, I tried using ModelComputeDtype, but it would randomly have memory problems for low Vram users. So I am just suggesting the command line method instead.
@luneva that's interesting. In my case, I can't even get this workflow to finish if I use the --fp32-unet command.
@Gh0sTt then may I suggest using --bf16-unet instead then if that helps. (or just stick with comfyUI 0.3.77) Which you can downgrade to if you revert to that TAG version in git.
v2.2 RELEASED! This will be my last big update for a while, until I find any new big image improvements. Hope you enjoy the new Detail vs Cinematic setting, it'll give you nice control!
V2.2 Added Detail vs Cinematic Settings section for maximum user control
Retuned for even more detail/realism.
Added Prompt Retry/Image Override Section
Cleaned Up Wires with Get/Set Nodes
Added FP32 FIX NOTES for Noise bug with ComfyUI Updates
(Impacting some users after 0.3.77+ Version, can now upgrade ComfyUI)
Cant wait to test this :)
@Lylyman be sure to share your pics, I can't wait to see what everyone comes up with with the new flexible settings!
should i add loras to both sampler 1 and sampler 2, do i need to adjust the value between this?
@dungtc01 Personally I add the Lora on the 2 samplers and I keep the same value. I don't know if it's the best method but it works pretty well!
@Lylyman it's oke? since the second sampler use 30 steps much more over 6 steps of the first
@dungtc01 no, the 2nd step doesn't use 30 steps, the 2nd sampler only uses about 6 steps. The first sampler's lora group creates the overall composition, and the 2nd sampler's lora group finishes with the details and realism level.
Adding the same lora to both lora groups would be fine. As long as you understand that when you apply a lora to the 2nd sampler, it may affect the level of realism on how it finishes the image, so you'll have to test whether your chosen Lora affects the realism negatively (unless that's what you want), and adjust the strengths of the lora.