Eric Qwen-Image Generate

Eric Qwen-Image Generate is the straightforward text-to-image node in this pack: prompt in, image out, single pass. No multi-stage upscaling, no ControlNet, no Spectrum - just the Qwen-Image-2512 model doing what it does best. If you want to see what the model itself can do before you bolt on all the acceleration, this is the node.

Before you get excited, the honest positioning from the KB: Qwen-Image is a heavy, admired generator, not the community's daily driver. It's a 20B MMDiT that takes LoRAs insanely well and has best-in-class text rendering (especially Chinese glyphs), but it is a VRAM hog and it is slow on consumer cards. One user on a 5090 reported the sibling UltraGen node taking over an hour for a 2 MP image before killing it. This node is for when quality and resolution beat speed - and for leaning into the model's two superpowers: legible text in the image, and long detailed prompts.

The inputs that matter

• prompt - required, and it deserves your attention. Qwen-Image uses Qwen2.5-VL as its text encoder, an LLM, so write sentences, not tag soup. Weighting syntax like (red:1.4) is passed through as literal punctuation and does nothing. The KB is unambiguous about this: prompt it like an instruction. Detailed natural-language descriptions are where the model shines.
• negative_prompt - optional; leave empty for the stock behavior, or use the official Qwen negative.
• aspect_ratio - preset dropdown (default 1:1 Square). This node's schema uses ratio + megapixels rather than width/height, so you pick a shape, not pixels.
• target_mp - target megapixels (default 4). The node computes exact width/height from ratio + MP, dimensions aligned to 16 px. Start lower for tests.
• steps - 50 for the base model (the default), fewer if you're using a step-reduction LoRA.
• true_cfg_scale - 4.0 default. This is true CFG: two full transformer forward passes per step (conditional + unconditional). Values above 1 enable the dual-pass; it's the honest version of CFG, and it's why these runs are slow.
• seed - 0 = random.

Output is image.

The version to use

The loader supports either Qwen-Image or the December 2025 Qwen-Image-2512 refresh, and 2512 is the one to point at: more realistic humans, finer texture, stronger text, and it ranked top of the open field on AI Arena at release. Same architecture, same loader, newer weights - no reason to use the original.

Practical notes

• It's a single clean pass - which is its virtue: no intermediate upscales to babysit. If you want the big multi-stage output, that's Multi-Stage Generate or UltraGen.
• VRAM math is brutal. The README's table (24 GB ≈ 2 MP, 48 GB ≈ 6 MP) applies. At 4 MP default on a 24 GB card you may be swapping; drop target_mp to 1-2 for iteration.
• LoRAs plug in upstream via Apply LoRA - and its stage weights don't matter here, since this is the single-pass node (stage-1 weight is the one that applies to non-UltraGen nodes, per the tooltip).

Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/, and load Qwen-Image-2512 with Eric Qwen-Image Loader first. The model download is the real install step - it's a full diffusers folder, not a single safetensors.