MOSS-Image ecosystem roundup: Spectrum speeds up to 12 steps on an RTX 3060
optimisticalish · reddit · 2026-10-19
A roundup of new MOSS-Image ecosystem tools and resources:
- Spectrum acceleration: plug in the 'auto' node after the model, no turbo LoRA needed—1168px (1.3Mpx) generations in as low as 12 steps / 30 seconds on an entry RTX 3060 12GB; ComfyUI's built-in Kitchen Attention has no effect on MOSS speed
- SFT DeBand (int8): models with the vertical color-stripe fix merged in, tested working
- BBox1000: resolution-independent 0–1000 BBox layout converter fixing earlier cross-resolution issues
- PromptWeight: the Qwen3.5 text encoder ignores ComfyUI's normal (word:1.2) weights, so this node applies them inside the diffusion model's attention, e.g. (red hat:-1)
- Fizgig 8.5: adds MOSS LoRA/Slider training, trainable on 12GB cards; author confirms resmultistep+simple beats euler and doesn't break Spectrum
- Style LoRAs: gouache and psychedelic style LoRAs; a claimed 4-step turbo LoRA (untested, first one over-simplified)
- Celebrity knowledge: initial list of known celebrities; no idea about Lovecraft or Tolkien
More from coding & agent
- The recipe behind Claude's all-code video: add cliptude MCP to generate hour-long videos — drrickio · 2026-10-19
- Claude made an entire video in pure code: 2 prompts, 13 scenes, 2,148 frames, no video model — drrickio · 2026-10-19
- "The agent says done — then my real job starts": devs compare verification workflows — falafel777 · 2026-10-19
- lexfridman-cn open sourced: one Python command builds 506 static episode pages — vista8 · 2026-10-19
- Lex Fridman Chinese companion site goes live alongside its open-source repo — vista8 · 2026-10-19
- All 506 Lex Fridman podcast episodes converted to bilingual reads with local ASR, now open source — vista8 · 2026-10-19