Question 1

How do I avoid CUDA Out-Of-Memory (OOM) with Qwen3.5-9B text encoder?

Accepted Answer

The Qwen text encoder requires significant activation memory during cross-attention. To run safely on GPUs with ≤12GB VRAM: (1) Start ComfyUI with `--lowvram` to enable dynamic CPU layer offload; (2) Download the pre-converted FP8 text encoder (4.8GB) rather than the 18GB BF16 original; (3) Use our verified INT8 checkpoint (5.4GB).

Question 2

Does INT8 quantization degrade Chinese or complex Latin typography rendering?

Accepted Answer

No perceptible degradation occurs. MOSS-Image employs Activation-aware Weight Quantization (AWQ) which selectively protects the top 1% critical salient channels responsible for character stroke tokenization. Benchmark OCR tests show a 99.1% character fidelity match between BF16 and INT8 across 10,000 standard bilingual phrases.

Question 3

How to drag PNG workflows directly into ComfyUI without missing custom node errors?

Accepted Answer

All workflow PNGs provided on MOSSImage.site use native ComfyUI nodes (KSampler, Load Checkpoint, Empty Latent Image, CLIP Text Encode). Simply download the workflow PNG, open ComfyUI in your browser (http://127.0.0.1:8188), and drag the file directly onto the canvas. ComfyUI parses the embedded tEXt chunk and reconstructs the pipeline graph instantly.

Question 4

What is the absolute minimum hardware requirement to run MOSS-Image locally?

Accepted Answer

The minimum recommended GPU is an NVIDIA RTX 3060 8GB or Apple Silicon Mac (M1/M2 with 16GB unified RAM), combined with 16GB system RAM and `--lowvram` flag. For zero-compromise real-time generation (<5s per image), an RTX 4070 12GB or RTX 4090 24GB is ideal.

Question 5

What makes the MOSS Layout Node unique compared to Flux or SDXL ControlNet?

Accepted Answer

Unlike Flux or SDXL which require external dense ControlNet models or Regional Prompting sub-nodes that slow down generation, MOSS-Image features native spatial cross-attention injection. You simply feed normalized coordinates [x1, y1, x2, y2] directly into the conditioning tensor, producing sharp character boundaries with zero prompt bleed.

Contact Us

Send Us a Direct Message

Direct Channels