⚗ AlchemicMoss — MOSS-Image workflow (prompt by its own Qwen3.5 encoder, framing presets, calibrated sampler)
2
10
2
⚗ AlchemicMoss for MOSS-Image
Write an idea, press Queue: MOSS-Image's own text encoder writes the prompt, the sampler is measured on your checkpoint.
⚗ Idea and direction: AlchemicaMente · code: Claude (Anthropic) · license: GPL-3.0
Custom nodes and documentation: https://github.com/Nessuno76/ComfyUI-AlchemicMoss
It is the MOSS-Image branch of ⚗ AlchemicaSampler (Krea 2), with its own node names: the two packs install side by side.
What it does
1. Prompt written by MOSS-Image's own encoder. Write an idea in any language: Qwen3.5-9B, the encoder MOSS-Image already uses, writes a scene, a detail and a negative. No Ollama, no external server: the nodes only talk to ComfyUI. Text to render goes in "double quotes": every word on these pictures was asked that way and came out right.
2. Framing and look presets with thumbnails: shot (close-up … full body … panoramic), camera angle, view, light, atmosphere, background.
3. The negative where it works. MOSS-Image is not distilled: the negative goes through a real cfg (2.5). When the prompt asks for a written text, text, letters, signature leave the negative, so it does not fight the lettering.
4. KSampler with a report. Any ComfyUI sampler and scheduler, or the Alchemica ones (alchemica_euler_2m, alchemica_calibrated): a sharper look on the same seed.
5. Calibration. Measures how your checkpoint builds the image and writes a profile; profile (auto) finds it. The profile of moss_image_sft_int8_convrot (12 prompts) is included.
Requirements
· ComfyUI 0.42+ (native MOSS-Image support) · custom nodes ComfyUI-AlchemicMoss 1.0.1 (ComfyUI-Manager, or the GitHub link)
· From Comfy-Org/MOSS-Image on Hugging Face: moss_image_sft_int8_convrot.safetensors (or bf16) → models/diffusionmodels · qwen3.5_9b_w4a8_gptq.safetensors (or int8convrot) → models/text_encoders · flux2-vae.safetensors → models/vae
· Tested on an RTX 5070 Ti 16 GB: about 21 s of sampling at 1.5 MP, 25 steps. The first run takes longer while the encoder writes the prompt.
How to use
1. Install the nodes, restart ComfyUI.
2. Open AlchemicMoss_T2I, write your idea in the Prompt node, press Queue.
3. For another checkpoint (bf16, a fine-tune, a LoRA): open AlchemicMoss_CALIBRATION, batch count 12, Queue. From 8 runs the profile is reliable.
Honest notes
· MOSS-Image needs a real cfg (2.5-3) and 25-50 steps: at cfg 1 the image is weak.
· Vertical stripes on skies and walls are a trait of the SFT checkpoint; a DeBand LoRA helps.
· Launch ComfyUI without--use-sage-attention: it washes MOSS-Image out.
· NAG, SEG and PAG of AlchemicaSampler are Krea 2 tools and are not used here.
License
Workflows and nodes: GPL-3.0. Models are not included: MOSS-Image (OpenMOSS) is Apache-2.0; for the Qwen3.5 text encoder see its model card.
Thank you
To OpenMOSS for MOSS-Image, Comfy-Org for the native support, the Qwen team, and everyone whose feedback on AlchemicaSampler asked for this.
⚗ — AlchemicaMente

