These workflows are optimized for NanoMix, but they work equally well with other Illustrious-based models such as Hassaku, Janku, Nova, and WAI — so feel free to use them with whatever checkpoint fits your style! They use RealESRGAN-2x for upscaling, available in the ComfyUI Manager or on OpenModelDB.
Single-Character Workflows:
The single-character workflows are built entirely with ComfyUI's native nodes — no custom nodes required. I’ve organized them into clean, minimal node groups to keep things simple, so you can focus on prompting without getting bogged down in technical details. They're ideal for single-character generations and straightforward compositions. Use Deathspike.Portrait for portraits and Deathspike.Wallpaper for wallpapers.
Multi-Character Workflows:
The multi-character workflows are designed for two or more characters, while keeping each character distinct and prompt-accurate. They use my Tiled Attention Couple custom node and are organized into minimal node groups to stay as clean and user-friendly as possible. You can start with a simple two-character setup, and then expand it to support three, four, or more. Use Deathspike.Portrait.Tiled for portraits and Deathspike.Wallpaper.Tiled for wallpapers.
Description
FAQ
Comments (13)
Sir, what a trick — the entire world is indebted to you
Sir, looks like there is a problem with the tokenizer, when you try to add promt descriptions with diferent token amouth.
This is the ouput:
Sizes of tensors must match except in dimension 0. Expected size 77 but got size 154 for tensor number 2 in the list.
Any idea of how to solve it,
I haven't encountered this error, except when I mismatch the Tiled Attention Couple resolution in a _Tiled_ workflow. Are you sure you haven't modified the resolution? If you're sure, which workflow gives the error, using which prompt? Please provide as much detail as you can, so I can reproduce the error on my end.
with this promt:
2girls, side-by-side, masterpiece, highres, very awa, best quality, sharp focus, traditional media, realistic, beautiful detailed, high contrast, portrait, Mature, grown old, BREAK delia ketchum \(pokemon\),brown hair, brown eyes, parted bangs, ponytail, low ponytail, huge breast, seductive smile, layered clothes, pink blouse, short sleeves, blouse, (blouse removed), exposing her bra, white front-hook bra, skirt, long skirt, blue skirt, (skirt lifted), white panties, skindentation, BREAK misty \(pokemon\), orange hair, short hair, side ponytail, aqua eyes, small breast, small waist, excessive pubic hair, fat labia, tongue out, yellow tank top, crop top, braless, (no bra, bare breast), lifting her shirt to reveal underboob, blush, black panties, (pulled down panties), denim shorts, (pulled down shorts), loose suspenders, hanging,
I solved it by removing duplicates and unnecessary tokens.
Thanks! I reproduced the error with your example prompt. The issue was in Tiled Attention Couple. If a tile had a different CLIP token length than the base, this error would occur. I pushed an update for the custom node, so once you update Tiled Attention Couple to v1.2.1, this prompt should work. It may take some time before the update shows up on Comfy Manager, if you use that.
Deathspike Wow Sir, thanks so much!!!! Im on it. Noisssse!!!
lovely workflow, just a question tho, maybe im missing something, but is there a way so that the images keep their metadata? the images that are generated and placed on the raw folder are striped of their metadata, which i would like to keep, or what adjustment would be necessary for that
Thank you! The Raw folder is used for output of intermediate steps and saving them uses the built-in SaveImage node that comes with ComfyUI. There is nothing in the built-in node that prevents the metadata from being saved. When I drag and drop one of them into my ComfyUI, I see the entire workflow as expected. Perhaps you changed the behavior of the built-in SaveImage through other custom nodes or configurations?
Deathspike thanks for answering, i should've been clearer, the image keeps general metadata like sampler, steps, cfg and seed, which is indeed metadata, but the prompt metadata is striped
ishidddd This does not happen for me. Even the output images for intermediate steps contain the entire workflow, including prompts. Since the workflow just uses the built-in SaveImage, yours should behave the same way. I suggest looking at the custom nodes you have installed. It's possible one of them money patches the built-in node and changes its behavior.
Deathspike
its alright, btw i also wanted to ask something else, is there a way to do something similar to normal so it can do img2img? or if it is already capable with another node, because using Load image lacks the node connection.
Would be helpful for more complicated gens
ishidddd Sorry, I don't have any img2img/inpainting workflows at this point in time.
Deathspike its alright no worries, just wondering about that