| backend | COMBO | Official PE (local Qwen3.5-9B) | 7 options: Official PE (local Qwen3.5-9B), Local LLaMA (GGUF+mmproj), Local HF (Transformers), API (OpenAI-compatible), BSAI Strata (Qwen3.8-Flash-Next @8080), BSAI Nemotron (llama.cpp @8085), +1 |
| prompt_text | STRING | A girl in a red dress standing under a cherry tree, holding bubble tea, evening light | — |
| system_template | COMBO | Qwen-Image-2.1 Official PE-T2I System Rules [official_t2i] | 354 options: No template (clear output, no system rules) [none], Qwen-Image-2.1 Official PE-T2I System Rules [official_t2i], Qwen-Image-2.1 Official PE-I2I System Rules [official_i2i], Official API prompt_extend Rewrite Style (with examples) [api_prompt_extend], Official Prompt Formula: Subject + Scene + Style + Lens + Mood + Details [formula_official], Official Poster / Slide Layout Template [layout_poster], +348 |
| custom_system_prompt | STRING | | — |
| user_requirement | STRING | | — |
| temperature | FLOAT | 0.700–2 | — |
| top_p | FLOAT | 0.950–1 | 核采样阈值。注意：ComfyUI v0.42+（#17221）修正了 top-p 边界语义（>= 判定 + 右移一位），同一 seed 的输出与旧核心不保证一致；本地 LLaMA (llama.cpp) 后端不受影响。 |
| top_k | INT | 400–1000 | — |
| repeat_penalty | FLOAT | 1.100–5 | — |
| seed | INT | 00–18446744073709550000 | — |
| max_tokens | INT | 409664–32768 | — |
| speed_preset | COMBO | Standard 25 steps/CFG1 (recommended) | 14 options: Standard 25 steps/CFG1 (recommended), Fast 15 steps/CFG1 (iteration preview), Turbo 10 steps/CFG1 (quick draft), High quality 35 steps/CFG1 (detailed output), Cached 20 steps/CFG1 (with EasyCache node), Viggle Turbo 6 steps/CFG1 (official node, recommended), +8 |
| pe_mode | COMBO | auto | 3 options: auto, T2I (text-to-image), I2I (image-to-image / edit) |
| min_p | FLOAT | 0.000–1 | — |
| thinking | BOOLEAN | true | — |
| mtp | COMBO | auto | 6 options: auto, off, 2, 3, 4, 5 |
| llm_model_name | COMBO | <No GGUF model found> | 1 options: <No GGUF model found> |
| mmproj_name | COMBO | | 1 options: |
| chat_handler | COMBO | auto (auto-detect Qwen3.8/Gemma4/GLM4.6 and other supported models from the model name) | 15 options: auto (auto-detect Qwen3.8/Gemma4/GLM4.6 and other supported models from the model name), MTMD generic multimodal (recommended: Qwen3.5+/Gemma4/GLM with mtmd-mmproj), Qwen 3.8 / 3.6 / 3.5 vision (requires updated llama-cpp-python), Qwen 3 / 3-VL vision (requires updated llama-cpp-python), Qwen 2.5-VL / 2-VL vision, Gemma 4 vision, +9 |
| n_ctx | INT | 8192512–65536 | — |
| n_gpu_layers | INT | -1-1–200 | — |
| load_mtp | BOOLEAN | false | MTP (Multi-Token Prediction) 层保留开关。官方 ComfyUI v0.42 / 新版 llama.cpp 已支持 Qwen 3.x/3.5 MTP（含图像场景，含 MTP embedding offload 修复）；模型带 nextn_predict 元数据且关闭本开关时自动剥离 MTP 层。追求吞吐加速可开启（需底层 llama.cpp 支持，开启失败时模型加载会报错，回退关闭即可）。 |
| n_cmoe | INT | 00–256 | Split MoE experts across CPU/RAM layers (-ncmoe). Only effective for MoE models, not dense.
Qwen-35B-A3B is most stable around 24; 0 = off. Current 0.3.36 does not support it yet; it applies automatically after upgrading. |
| kv_cache_quant | COMBO | auto (f16, no quant) | KV cache quantization (-ctk/-ctv, q4_0 = video effect). Lower VRAM and faster, at a slight quality cost. |
| use_mmap | BOOLEAN | true | use_mmap; can disable when RAM is tight / swapping (--no-mmap). |
| use_mlock | BOOLEAN | false | use_mlock (--mlock) locks the model in RAM to avoid page-out. |
| flash_attn | COMBO | auto | -fa. auto = enable when the model supports it. |
| threads | INT | 00–256 | -t thread count (video example: 14); 0 = auto. |
| keep_loaded | BOOLEAN | false | — |
| api_base | STRING | https://dashscope.aliyuncs.com/compatible-mode/v1 | — |
| api_key | STRING | | — |
| api_model_name | STRING | qwen3.5-vl-9b | — |
| timeout | INT | 12010–600 | — |
| strata_quant | COMBO | Auto (auto-adapt) | Strata quantization. Auto picks based on VRAM/RAM. |
| ninfer_model | COMBO | (Auto — first model found) | NInfer model to load (models/LLM/*.ninfer). Changing it restarts the NInfer engine on the next run. |
| nemotron_model | COMBO | (Auto — first model found) | Nemotron GGUF model to load (models/LLM/Nemotron/*.gguf, excludes mmproj). Changing it restarts the llama.cpp server on the next run. |
| hf_model_name | COMBO | | 1 options: <No HF model found> |
| hf_task | COMBO | <MORE_DETAILED_CAPTION> | 5 options: <MORE_DETAILED_CAPTION>, <DETAILED_CAPTION>, <CAPTION>, <OD>, <REGION_CAPTION> |
| hf_device | COMBO | auto | 3 options: auto, cuda, cpu |
| hf_keep_loaded | BOOLEAN | true | — |
| preview_only | BOOLEAN | false | — |
| clipopt | CLIP | | — |
| image_1opt | IMAGE | | — |
| image_2opt | IMAGE | | — |
| image_3opt | IMAGE | | — |
| image_4opt | IMAGE | | — |
| image_5opt | IMAGE | | — |
| image_6opt | IMAGE | | — |
| image_7opt | IMAGE | | — |
| image_8opt | IMAGE | | — |
| image_9opt | IMAGE | | — |
| image_10opt | IMAGE | | — |
| image_11opt | IMAGE | | — |
| image_12opt | IMAGE | | — |
| image_13opt | IMAGE | | — |
| image_14opt | IMAGE | | — |
| image_15opt | IMAGE | | — |
| image_16opt | IMAGE | | — |
