⚗ AlchemicMoss
The AlchemicaMente line for MOSS-Image: a prompt written by MOSS-Image's own Qwen3.5 encoder, framing and look presets, a KSampler with schedules measured on your checkpoint, and the Calibration that measures it. Installs next to ComfyUI-AlchemicaSampler (Krea 2). Idea: AlchemicaMente · code: Claude (Anthropic).
⚗ AlchemicMoss
The AlchemicaMente line for MOSS-Image (OpenMOSS, Apache-2.0), natively supported by ComfyUI. Four nodes: Load → Prompt → KSampler → Save, plus the Calibration that measures your checkpoint.
It is the MOSS-Image branch of ComfyUI-AlchemicaSampler (Krea 2), with its own node names: the two packs install side by side.
Idea & direction: AlchemicaMente · code: Claude (Anthropic) · GPL-3.0
Install
ComfyUI Manager → search AlchemicMoss, or:
cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicMoss
No extra Python packages. ComfyUI 0.42 or newer (native MOSS-Image support).
| File (from Comfy-Org/MOSS-Image) | Folder |
|---|---|
| moss_image_sft_int8_convrot.safetensors (or _bf16) | models/diffusion_models |
| qwen3.5_9b_int8_convrot.safetensors or qwen3.5_9b_w4a8_gptq.safetensors | models/text_encoders |
| flux2-vae.safetensors | models/vae |
Then open the workflow AlchemicMoss_T2I (Workflow → Browse templates → AlchemicMoss), write an idea, press Queue.
The nodes
| Node | What it does |
|---|---|
| ⚗ AlchemicMoss Load | model + Qwen3.5 text encoder + flux2 VAE. The encoder is always loaded with the moss_image type, the most common setup error. A Krea 2 model stops with a clear message. |
| ⚗ AlchemicMoss Prompt | an idea in any language → MOSS-Image's own Qwen3.5 encoder writes the prompt (or as written keeps your text). Lists, wildcards, examples. At the bottom: shot / camera angle / view / light / atmosphere / background presets with thumbnails (📷). It makes the start latent in MOSS-Image's format (128 channels, 1/16) and a negative. |
| ⚗ AlchemicMoss KSampler | any ComfyUI sampler and scheduler, plus the Alchemica ones (alchemica_euler_2m, alchemica_calibrated, alchemica_dual, flux_shift). A report says what it did. |
| ⚗ AlchemicMoss Calibration | measures how your checkpoint builds the image and saves a profile. The KSampler's profile (auto) finds it by itself. |
What to know about MOSS-Image
- Not distilled: it needs a real cfg (2.5-3) with the Prompt's negative connected, and 25-50 steps. The defaults:
er_sde+simple, 25 steps, cfg 2.5, one stage, 1.5 MP. - Prompts: longer prompts suit it (
words150). Text to render goes in "double quotes": the Prompt tells the encoder so, and dropstext,letters,signaturefrom the negative when the prompt asks for a written text. - Framing presets go at the head of the prompt. The composition phase of AlchemicaSampler (framing text first, full text later) stays off: on MOSS-Image the framing text alone would decide clothes and colours.
- NAG / SEG / PAG are Krea 2 tools: here the negative works through the cfg.
- Vertical stripes on skies and walls are a trait of the SFT checkpoint. A DeBand LoRA helps.
--use-sage-attentionwashes MOSS-Image out. Launch ComfyUI without it.- Speed on an RTX 5070 Ti 16 GB, int8 model + w4a8 encoder, 1.5 MP: about 21 s of sampling for 25 steps.
Calibration
A profile of moss_image_sft_int8_convrot (12 prompts, cfg 2.5) ships in profili/. For another file (bf16, a
fine-tune, a LoRA stack) open AlchemicMoss_CALIBRATION: it measures one prompt of a list per run, with the
Prompt's negative and cfg 2.5. Queue it 8 or more times (Queue with a batch count of 12): from 8 samples the
profile is reliable and alchemica_calibrated and every -1 value of the KSampler come from it.
Measured on moss_image_sft_int8_convrot: composition settled at sigma 0.93 → 0.80, detail window 0.71 → 0.25.
Workflows
| Workflow | | |---|---| | AlchemicMoss_T2I | Load → Prompt → KSampler → Save: write an idea, Queue | | AlchemicMoss_CALIBRATION | measures your checkpoint, one prompt per run |
License
GPL-3.0, the same license as ComfyUI. See LICENSE, CREDITS.md and THIRD_PARTY_NOTICES.md.