Workflow made in response to one comment in this article.
This workflow for ComfyUI uses the 3-pass approach to convert images from Anime style checkpoints (NoobAI Anime) to photorealistic render.
Main usecase is to get a composition, character, dress, etc. from anime style checkpoint and make the same image in photorealistic style. This approach is useful to prepare photo-realistic training dataset for anime checkpoint.
Read this article for more details about the workflow.
Description
It takes random image from specified folder, caption it with WD1.4 tags and Qwen VL 2.5, then applies 2-pass realistic conversion. It uses non-Core nodes (additional nodes installation is required).
FAQ
Comments (19)
I've been hoping for such a workflow for a while now, as Forge has too many limitations in this regard. It's time to get into ComfyUI, and not just to test WAN or other things.
I'm not using A1111 for so long time now... not needed. Even Inpainting is good now in ComfyUI.
I was able to run the 3-pass workflow, but not the 2-pass one, I have an error at the beginning : 'float' object cannot be interpreted as an integer" on the Random Number Node. I gave a correct directory path with 3 pictures in it.
"'float' object cannot be interpreted as an integer"
On which node? Probably you have deleted the conversion of Float number into Int index somewhere.
ComfyUI must highlight the error source with purple.
@homoludens On the "random number" node. I've re-downloaded an launched the workflow, only changed the directory path.
I also have no idea of what to put for "CinEro-V6-R4e_00004_.lora_2025-08-24_16-36-30-save-53712-144-0.safetensors"
@Brewce I forgot to cleanup things. It is experimental lora of R5 release. Just delete this node.
>> On the "random number" node
This node should have a connection at "maximum" field.
This field should be connected with output from "IntToFloat" node (ComfyMath) package.
@homoludens they are connected, i didn't touch the workflow. I don't know how to see what is the IntToFloat output.
I gogled exact your error and found the issue on github
https://github.com/heheok/comfyui_wan2.1_vace_infinite_helpers/issues/1
But cannot understand the proposed solution and what is the root cause
Can't open this node in VS Code right now. need to sleep a bit...
If you know the number of images in folder (Ctrl+A then properties) then you can remove the connection and write the exact value by hands. It might do the trick (50/50)
@homoludens I found it ! On the Random Numbe rnode, I just set "number_type" to float... it's set to integer by default.
cinEro-v6-re4_0004_.lora where to find this?
It is integrated into CinEro ILL v6 R4e release most likely.
Could you please give a link on the image where you found this - I need to check the workflow and give you info if that lora was DEactivated at render time.
Anyway, this particular LORA does nothing for the matter of the workflow itself.
It just an intermediate resource in CinEro training process.
I've released a new similar LORA which might be more interesting for you.
You have to t2i first?, you can't just upload an image?
You can replace the t2i anime-ish pass with just an image, but this defeats the main reason I've created this workflow.
The idea is the get an image from the anime-ish checkpoint for the given text prompt and then produce the image with photorealistic colors, lighting, textures, but having the very close scene composition and narrative. This will allow to train a LORA that don't affect composition but converts colors and anatomy into realism.
Hope this make sense for you.
For those who just wants to get a realistic image and post it to the community it is simpler to replace the 1st pass with an Image and Text inputs.