Anima → Krea2 → SeedVR2 | Three-Stage Anime Pipeline (txt2img → repaint → upscale)
A single-queue workflow that combines Anima's tag-based anime generation with Krea2 Turbo's natural-language refinement, finished with SeedVR2 upscaling.
What it does
Anima base (txt2img) — Anima base v10 generates the composition from Danbooru-style tag/nature language prompts. Full LoRA support via Power Lora Loader.
Krea2 repaint (img2img) — The base image is passed straight into Krea2 Turbo through the shared
qwen_image_vae(no quality loss from format round-trips). 8 steps / CFG 1, denoise 0.3 by default:0.2–0.35 → light polish, keeps Anima's composition and linework
0.4–0.55 → stronger Krea2 restyling, more detail repainted
SeedVR2 upscale — lanczos ×0.7 pre-scale, then SeedVR2 7B upscales to 2000px.
Features
Separate LoRA loaders and prompts per model — tag prompts for Anima, natural language for Krea2. Describe the same scene on both sides for best consistency.
Shared seed — one Seed Generator feeds both KSamplers, so one roll syncs the whole chain. Set it to
fixedto reproduce and iterate.One-click upscale bypass — Fast Groups Bypasser toggles the SeedVR2 group for fast preview rounds (saves VRAM too).
Built-in A/B comparer — slide between the Anima base and Krea2 repaint.
Intermediate saves at every stage (base / repaint / final), core SaveImage nodes only — workflow embeds in PNG metadata as usual.
Required custom nodes
rgthree-comfy: Power Lora Loader, Image Comparer, Fast Groups Bypasser
ComfyUI-SeedVR2_VideoUpscaler: SeedVR2 DiT/VAE loaders + upscaler
ConditioningKrea2Rebalance: Krea2 NSFW
Description
Comments (2)
I was intrigued because I've never really worked with Krea2 before but I'm not really sure what the Krea step is supposed to do. If I match my prompt, it is a pretty close approximation of my images but the results are a little muddier, the lineart is more jagged and it seems to like the not connected lines lineart style (like old hand painted flash animations), and if I do a fully custom natural language interpretation of the prompt for the krea step, I get about the same results. Then the upscale just magnifies the rougher art, it almost seems like it would make more sense to go straight from anima to seedvr. I'm clearly missing something! I tried aiming for stylistic changes too as like a remix but it never really produced the right results. I'll have to play with it some more..
Thanks for testing it! You are not necessarily missing anything. The Krea2 step is not intended as a universal improvement or mandatory step before SeedVR2.
The motivation is that Anima is a very nice base model for style and composition, but in my tests it can struggle with body understanding, hands/fingers, limbs, and more complex anatomy. Krea2 seems stronger with human/body details and background interpretation, so I use it as a low-denoise repaint/correction pass.
That said, if the Anima output is already clean, especially for simpler or lineart-heavy images, Anima → SeedVR2 or SD Upscale may absolutely be better. If Krea2 makes things muddier or adds jagged/disconnected lineart, I would lower the denoise or bypass it. The intended use case is more “fix/polish details while preserving Anima’s composition,” not full remixing.