AI · Tip
Wan 2.2 NSFW Video Generation Recipe
Specific sampler settings, LoRA configurations, and prompt workflows for Wan 2.2.
This note details a battle-tested ComfyUI setup for generating NSFW video using the Wan 2.2 model. It shares exact scheduler and sampler combinations, LoRA weights, and a SillyTavern prompt-crafting workflow to achieve stable outputs. It is particularly useful for creators looking to bypass trial-and-error render tests on cloud GPUs.
How it works
Final result settings:
- seed - 938315953321073
- steps - 4
- cfg - 1.0
- sampler - sa_solver
- scheduler - beta
- denoise - 1
- shift - 8
- CLIP Skip - 2 (well, -2 because ComfyUI)
- Coupled with the I2V FusionX LoRA at 0.5 and a RealismBoost LoRA weighted at 1.5
Workflow notes:
- Iterated by adding different LoRAs, adjusting shift settings, adding/removing tags to hone in on test 41; upscaling raw output beyond native quality made pictures worse.
- Frontend: SillyTavern, with a system prompt written by ChatGPT and refined by the author to craft prompts.
- SillyTavern's 'data bank' feature is basically RAG: the author indexed 20 previously-written prompts they liked and uses them for reference.
- Resolution varies by need: 360p for testing prompts, 720p for final videos.
- Total steps 30-35. High noise: sampler dpmpp_3m, scheduler beta, 18 steps. Low noise: sampler deis_4m, scheduler beta, remaining steps; in the last 3 steps swap the sampler to res_4s_krogstad for refinement.
- Runs on an H200 on RunPod using the author's Wan template.
- A 3-second video takes ~10 minutes to render at higher resolution, ~5 minutes at lower resolution.
- LoRAs: a variety, but mostly Instagirls; also mentions CivitAI LoRAs like Instagirls and FusionX (from tweet text).
What it's for
- Configure ComfyUI sampler and scheduler schedules for Wan 2.2 video generation
- Pair image-to-video LoRAs like FusionX and RealismBoost at calibrated weights
- Use SillyTavern's data bank as a prompt-referencing RAG tool for video prompts
Key points
- Dual-stage sampling schedule using dpmpp_3m for high noise and deis_4m/res_4s_krogstad for low noise refinement
- LoRA weighting recommendations including I2V FusionX at 0.5 and RealismBoost at 1.5
- ComfyUI parameter set covering CFG, shift, CLIP skip, and sa_solver steps
- SillyTavern RAG setup indexing tested prompt banks for prompt drafting
- Hardware and resolution benchmarks for 360p draft versus 720p final runs on RunPod H200s
Pricing
Free
Details
- Type
- Tip
- Category
- AI
- Topic
- AI Video Generation
- Open source
- No
- Platforms
- Self-hosted
- For
- AI video creators, ComfyUI power users, and digital artists working with Wan 2.2
- Tags
- wan2.2, comfyui, ai video, lora, runpod, sillytavern
Why it was saved
- @deedydas: Concrete, reproducible parameters someone actually used to get high-quality results — the kind of settings hard to find without scouring communities like r/unstable_diffusion.
Related
AI creative suite for generating and editing video, images, and animation
AI platform generating and editing video, image, and synchronized audio.
AI video and image generator known for cinematic motion and physics
AI-assisted multitrack video editor and text-to-video generator for creators and teams.
AI video and image generator with advanced motion control and physical simulation
Generative AI video model producing cinematic footage with native audio
Updated