Loop ForgeSmaller text encoders: the cafe sign
From MOSS-Image add-ons, tested locally
- Model
- MOSS-Image Base (moss_image_base_int8_convrot.safetensors)
- Output
- 1024×1024
- Sampling
- 25 steps, CFG 2.5, er_sde / simple
- Seed
- 101
- 9B encoder
- 42.0 s, peak VRAM 10.2 GiB
- Lite TE 0.8B
- 41.2 s, peak VRAM 7.6 GiB
- Mini 0.8B
- 41.0 s, peak VRAM 6.6 GiB
- Mini 4B
- 41.2 s, peak VRAM 8.8 GiB
- Ran on
- RTX 3060, 12GB (local)
With Mini 0.8B in place of the default encoder, the peak memory dropped by about a third without slowing anything down. The poster stayed readable right down to the small print at the bottom, so this is the first change I'd make.
Photograph of a man with short black hair, a trimmed beard and a white linen shirt with rolled sleeves, standing in front of a small corner cafe at golden hour. Above the door a hand-painted wooden sign reads "THE SLOW KETTLE" and a chalkboard beside him reads "Fresh bread at 7. Open late on Fridays." Warm light, shallow depth of field, 35mm photo.
nsfw, worst quality, bad anatomy, watermark, jpeg artifact
Notes
The official Base template. The 9B encoder is the w4a8 file the templates load. Lite TE is a Qwen3.5-0.8B with a trained head (MOSS Lite TE Loader); Mini 0.8B and Mini 4B are smaller Qwen3.5 models with a projection adapter (Load CLIP, MOSS + projection). Each feeds both prompts in place of the 9B encoder; nothing else changes.
Peak VRAM is the whole card as nvidia-smi reports it, minus what the desktop used with ComfyUI stopped.
cyberlangke/moss-image-mini-encoder on Hugging Face
moss-image-mini-encoder node on GitHub
cicalooo/MOSSLiteTE on Hugging Face
Comfy-Org workflow templates (the MOSS-Image templates ship with ComfyUI)
Run it yourself
This test loads no input files. Put the models below in place and drag the workflow onto ComfyUI.
| File | Put it in | Size |
|---|---|---|
| moss_image_base_int8_convrot.safetensors | models/diffusion_models | 5.4 GB |
| qwen3.5_9b_w4a8_gptq.safetensors | models/text_encoders | 7.3 GB |
| flux2-vae_small.safetensors | models/vae | 249.5 MB |
| mosslite_qwen3.5_0.8b_bf16.safetensors | models/text_encoders | 1.5 GB |
| mosslite_te_head_bf16.safetensors | models/mosslite | 316.9 MB |
| mini_qwen3.5-0.8b-w4a8.safetensors (published as weights/qwen3.5-0.8b-w4a8.safetensors, renamed) | models/text_encoders | 1.0 GB |
| qwen3.5-0.8b_w4a8_linear.safetensors | models/moss_encoder_adapters | 50.4 MB |
| mini_qwen3.5-4b-w4a8.safetensors (published as weights/qwen3.5-4b-w4a8.safetensors, renamed) | models/text_encoders | 4.0 GB |
| qwen3.5-4b_w4a8_linear.safetensors | models/moss_encoder_adapters | 126.0 MB |
Custom nodes: ComfyUI-MOSSLiteTE, moss-image-mini-encoder version 0.1.0. Everything else is part of ComfyUI 0.43.1 (frontend 1.57.6, templates 0.11.100).