aigen/docs/qwen21.md

Qwen Image 2.1 (engine qwen21) — host weights + Comfy graph

Host paths (this RTX 5080 box)

Role	Resolved path
ComfyUI root (Klein / host agent ComfyUI (1))	C:\Users\ianjm\AppData\Local\Comfy-Desktop\ComfyUI-Installs\ComfyUI (1)\ComfyUI
Models root (Desktop Shared via shared_model_paths.yaml)	C:\Users\ianjm\AppData\Local\Comfy-Desktop\ComfyUI-Shared\models

Weights must land under Shared so every Desktop instance sees them:

• diffusion_models\qwen-image-2.1-Q8_0.gguf (~7.59 GiB) — from abenzerps/Qwen-Image-2.1-Uncensored-GGUF
• text_encoders\qwen3vl_8b_int8_convrot.safetensors — from Comfy-Org/Qwen-Image-2.1 (INT8; BF16 8B VL OOMs on 16 GB)
• vae\qwen_image_2.1_vae_bf16.safetensors — from Comfy-Org/Qwen-Image-2.1 (not the old Qwen-Image 1.0 VAE)

Custom node: custom_nodes\ComfyUI-GGUF (UnetLoaderGGUF). Do not install a second GGUF pack.

Setup

powershell -ExecutionPolicy Bypass -File scripts\setup-qwen21.ps1

DiT download uses the exact host invocation:

hf download hf://abenzerps/Qwen-Image-2.1-Uncensored-GGUF/qwen-image-2.1-Q8_0.gguf

That lands in the Hugging Face hub cache; the setup script copies it into Shared diffusion_models. The abenzerps GGUF ships with kv_count=0 (no general.architecture) and Q8_0-quantized 1D RMSNorm weights (logical 128 → packed 136), which breaks Comfy’s fused rms_rope. scripts/tag-qwen21-gguf.py rewrites the file with general.architecture=qwen_image and promotes small/1D tensors to F32 so city96 UnetLoaderGGUF can load it.

Restart Comfy only when idle (COMFY_CONTROL_URL/status → gpu.busy=false). Confirm object_info lists UnetLoaderGGUF and the three filenames.

App

• Engine key: qwen21 · UI label: Qwen 2.1
• Generate (T2I) and Edit (same checkpoint, second graph). Compose / Iterate / Video / Extend / Music stay disabled unless a graph exists.
• Edit requires Start still → images.image_1; prompt slots use <image1> / <image2>. If the user omits <image1>, the runner prepends Keep the subject in <image1>.
• Face / outfit locks stay Klein semantics — they do not drive Qwen slots.
• Sampler defaults: euler / simple / cfg 1 / steps 25. Edit uses QwenImage21Cache (device auto, dtype int8). T2I uses ModelSamplingAuraFlow shift 3.1.
• Default Generate canvas follows the bench Aspect control on the Qwen-safe ~1 MP grid (EmptyLatentImage): 1:1 → 1024×1024, 16:9 → 1536×864, 9:16 → 864×1536 (and the other table rows). PE wh_ratio is advisory only and never sizes the canvas. Do not use native 2K bins on 16 GB.
• Edit follows image_1 via the encode node’s latent (resolution long-edge ~1024). Do not inject a picker EmptyLatentImage onto the edit sampler.
• Optional Enhance prompt (off by default): runs a separate PE-only Comfy graph (CLIPLoader + rewrite node), then frees VRAM and queues the existing T2I/Edit graph with the rewritten prompt. Never loads PE CLIP + DiT together on 16 GB. Fail closed if parse_ok is false or the rewrite is empty.
• PE weights (int8 only): text_encoders\qwen3.5_9b_qwen_image_2.1_pe_{t2i,i2i}.int8_convrot.safetensors — do not replace the image TE qwen3vl_8b_int8_convrot.safetensors.
• Optional Turbo (off by default): Viggle DMD LoRA Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors on the same Q8 GGUF via ViggleTurboLora + ViggleTurboSigmas (1.0, 0.9375, 0.875, 0.75, 0.5, 0.25), 6 steps, CFG 1, empty negative. Not a new engine. Enhance prompt stays compatible and is recommended with Turbo.
• Custom node: ComfyUI-Viggle-Turbo (viggle_turbo.py). Do not merge the LoRA with stock LoraLoaderModelOnly (lossy on int8/bf16).

Graphs: server/assets/studio2_qwen21_t2i.json, server/assets/studio2_qwen21_edit.json, server/assets/studio2_qwen21_t2i_turbo.json, server/assets/studio2_qwen21_edit_turbo.json, server/assets/studio2_qwen21_pe_t2i.json, server/assets/studio2_qwen21_pe_edit.json.