AI Workflow Any ComfyUI IMG2IMG workflow/tutorial?

J

JonoFret

Lurker
Joined
Apr 28, 2026
Posts
1
Reaction score
0
Does anyone have a reliable ComfyUI img2img workflow or tutorial they recommend? I am trying to refine my process and would love to see what is working well for others right now. Please drop a link or share your setup if you are open to it.
 
Klein 9B Distilled (edit or original t2i, worth getting both for learning in general), For the workflow, simply search for “Klein” in the ComfyUI template/community section. To load the model, use the GGUF version — it’s much lighter on VRAM. There’s currently only the BigLove Klein variant available. You can find the files by searching for user Granddyser on Hugging Face.
Also, the quantized Gemma 4 E4B models, when used through a custom LLM node, are extremely lightweight and responsive — especially the abliterated variants. They integrate smoothly into your generation flow, allowing you to run a preprocessing step to refine or generate prompts on the fly if you like that workflow.
Unfortunately, you’ll mostly need to go to Civitai for LoRAs, as Hugging Face archives for these specific models exist but the site has poor search functionality. Civitai can be a bit of a wildwest full of ego-driven posts, but the quality and compatibility of Klein LoRAs vary. That said, if you’re using the BigLove variant, you’ll already have a lot of flexibility even without LoRAs. When you do use them, you usually won’t need more than one or two.
For reference I use a RTX 3090.
 
Hey guys, idk if this is the right place to ask this
but im extremely new to comfyui and basically know nothing about it so sorry for the dumb question
So like my issue is that I just freshly downloaded comfyui and my software isnt ditecting my checkpoints even tho its in the model/checkpoints file and the models file obviously ends with .safetensors
I tried reloading comfyui and restarting everything but it just refuses to notice it. For instance Im trying to run cyberrealisticpony and ponyrealism.
Hope someone can help, thanks!
 
If you use the Desktop app via ComfyUI official website, then the installer would have asked to set a location for the software, it would end up being here:
📁 YOUR_INSTALL_PATH/
└── 📁 ComfyUI/
───└──📁 models/
──────├──📁 checkpoints/ <-- Drop your .safetensors checkpoints here
──────├──📁 loras/ <-- Drop LoRAs here
──────├── 📁 vae/ <-- Drop VAEs here
If you installed it manually using your own venv and terminal, then it will be similar, basically the same folder you git cloned the repo with the above structure
You can simply press the R key without needing to restart ComfyUI for the lists to refresh
 
Thank you so much! It was actually in my documents comfyui folder and not the one in C drive.
I was wondering do you have any simple to use SDXL based image2image workflow? Which basically behaves like grok or geminis image gen? Like usually on regular i2i the model just changes very specific details while keeping the original image intact. But I want it to basically take the "identity" as in the facial details body shape and clothes of a refrence image then prompt it into a totally different action than it is in the refrence image. For instance an anime character just standing refrence image with white background right? I want it to be a closeup portrait with a sea beach background and night time moonlight. Overall the whole logic is much like how you just ask grok or gemini to create something different out of your existing image with natural prompting.
For refrence, Im trying to do all this using Illustrious XL gguf and my card is 5060 8gb
Sorry for the long texts, Im very new to all this.
 
with a 5060 8gb, you could probably do Klein 4b edit without any issues, or even 9b if using gguf versions of the model.
4b official workflow is here, https://docs.comfy.org/tutorials/flux/flux-2-klein#flux-2-klein-4b-workflows
Klein prompting is different to others, it will use natural flowing language and not keyword tagging. Assume you add a reference you can start off small and just say things like, "Change the woman's clothing to white pleated skirt and red leather jacket", in more advanced cases you add more reference images and use specific prompting to guide the model to give you want you want.
It doesn't stop at that either besides face, clothing etc.. you can change entire lighting environments, have sun rays casting through a window or direction, and change day to night and so forth.
With smart prompting and klein you won't even need loras, only in edge cases.
 
To be honest with you, I need models and workflow for it that allow nsfw and can still handle 3D render type photoshoot output using 2D refrence images. Pretty sure flux models aren't cut for that nsfw stuff right?
 
Do you know of any adjustments to this workflow to run on Apple Silicon?
BF16 or FP16 would be great for bigLove or a LoRA format instead of the full model.
 
amazing information. this method is so much better than ezcreate or any of those online stuff. thank you boss :pepoLove:. are there any forums with up to date guides on how to learn more about image to vid generation or vid to vid generation besides here? I don't think the guides here for those are very up to date. WAN 2.2 or LTX 2.3? :pauseChampRev:
 
Sound > No Sound, Wan is also meh (only use it if you are desperate) so If you can run LTX 2.3, use that, either with heavy use of Loras on the original Lightricks model, or use the Sulphur and 10eros finetunes.
https://huggingface.co/TenStrip/LTX2.3-10Eros
https://huggingface.co/TenStrip/LTX2.3-10Eros_Workflows
https://huggingface.co/SulphurAI/Sulphur-2-base
Nothing in the way of easy guides, it'll get more complicated when you do video. This is the kind of stuff you can't have one button does all, however it isn't a censored site for paypigs like any of those online services you see... I scoff at people who think Grok is good, most of it is jank softcore and heavily refused prompts while wasting tokens.
 
seems like you got it worked out, mine was spitting out absolutely neon sludge images, would you kindly share your workflow please? because im exactly looking for ezcreates replacement now
 
For Flux 2 Klein 9B
CFG 1.0 for Klein 9B Distilled (this is typically that doesn't say base in the name and takes 4 steps for normal inference)
CFG 4.0~ for Klein 9B Base (this one is the non distilled varient and requires around 20 steps for normal inference)
The images you describe is a mix-up regarding the above.
If you use CFG 4 with a Klein Distilled model you get just that. So double check what you are using, and how. If you need starting platform workflow the ComfyUI official documentation website has them.
 
Im running 4B fp8, any idea how to enable nsfw on it like adding detailed vagina or anus? it basically smudges those places out. cant find loras for 4B
 
I wanted to add to the LTX vs WAN discussion.
WAN - Good prompt adherence, physics and motion especially for NSFW. Huge ecosystem with many models/loras/workflows to choose from. No sound built in, would need a separate workflow to add it in.
LTX - Frustrating Prompt adherence(LLM enhancement almost required), lackluster physics and motion. Has great sound, it is mostly trained as a 'talking head' model. It does have great system performance for the quality you get. Newer model, smaller ecosystem.
Overall i think the place you want to be is LTX especially as the ecosystem grows. There are workarounds for the crappy prompt adherence and we should be getting a new version soon hopefully.
 
There is a year of maturity on WAN 2.2 and more if you count WAN 2.1, and then you have Bernini R as of late last month which hybridizes WAN 2.2 a bit. All from the same source architecture.
LTX by way of design especially in 2.3, prefers loaded prompting due to its text encoder and how it was trained, and most have leant on using an LLM to prompt for them, which I do aswell, I had claude code a node just for this to ease the burden of vram use but also use llama cpp locally rather than the god awful ollama vision support (as using image references is good just from a feature standpoint). However there are probably 100 ways now to use an LLM via comfy whether it is in node or via local api, or if one had cash to burn god forbid use external SaaS.
In essense what I am saying is, we have access to both, best to just learn both. Since they both run in the 24GB vram budget, or less (undesirably)
 
never understand the hype surrounding ltx, for i2i qwen rapid aio is miles better
 
tried it and didn't like it at all. low quality images, too slow and it completely alters the faces of your "input" . not worth it imo.
 
You must log in or register to reply here.
Back
Top