Eric Qwen-Image Load Model

Eric Qwen-Image Loader is the entry point for the entire generation half of this pack. It loads the Qwen-Image (or Qwen-Image-2512) text-to-image pipeline from a local directory and hands you a single QWEN_IMAGE_PIPELINE wire. From there everything flows: Generate for a single pass, Multi-Stage and UltraGen for progressive upscaling, Apply LoRA to stack adapters, Spectrum to speed it up, ControlNet Loader + UltraGen CN for structure-guided work, Harmonize for composites.

Like its edit sibling, this is not a ComfyUI-native loader. It runs the real Hugging Face diffusers pipeline through from_pretrained, which is the pack's whole design bet: the official FlowMatchEulerDiscreteScheduler with automatic resolution-aware time shifting, instead of the generic UNET → KSampler abstraction where you'd hand-tune "Aura Flow Shift" values yourself. You don't need any shift nodes with this pack - the model's own config drives it. The trade is that the generic "Load Diffusion Model / Load CLIP / Load VAE" nodes produce wrappers this pipeline can't use, so you must load through here (or the Component Loader).

What you actually set

• model_path - path to a full Qwen-Image or Qwen-Image-2512 model directory. Needs the whole diffusers folder: model_index.json, scheduler config, tokenizer, and the sharded weights (~50 GB). Use 2512 - it's the December 2025 refresh (more realistic humans, finer texture, better text) and ranked top of the open field on AI Arena at release. No GGUF shortcut; this loader is full precision.
• precision - bf16 default, the right choice on RTX 40/50.
• keep_in_vram (True) - cache between runs so you don't re-read 50 GB each time.
• offload_vae (False) - VAE to CPU during transformer inference, saves ~1 GB.
• attention_slicing (False) - lower peak VRAM, slightly slower.
• sequential_offload (False) - aggressive CPU offload, very slow, but it's the difference between running and OOM on small cards. One user on an 8 GB 4060 found it mandatory with this pack's edit side; expect the same on generation.
• vae_tile_blend - cosine (default) or linear seam blending for tiled VAE decode. Cosine kills faint grid lines on gradients.

Output: pipeline (QWEN_IMAGE_PIPELINE).

The honest hardware talk

Qwen-Image is a 20B MMDiT and the KB's community signal is blunt about it: it's a VRAM hog and not a daily driver on consumer cards. The README's guidance for the edit model applies here too - 24 GB ≈ 2 MP, 48 GB ≈ 6 MP. And the speed is where people actually get burned: this pack's multi-stage generation nodes are heavy, and a 5090 user reported an hour-plus for a 2 MP UltraGen run before killing it. The author runs an RTX 6000 Pro. If you're on a 24 GB card, plan for single-pass Generate at low target_mp or the lighter Multi-Stage settings, not 30 MP marathons.

Two pack-wide gotchas. The requirements pin diffusers<0.38 and transformers<5 (diffusers 0.37 still imports HybridCache, which transformers 5.x removed - the comment in requirements.txt is worth reading); another pack forcing a newer diffusers can break this loader. And this pack is dual-licensed CC BY-NC 4.0 - fine for personal use, a real conversation before commercial work.

Install

ComfyUI Manager search "Eric Qwen Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

Restart, then the actual install step: download Qwen-Image-2512 from HuggingFace to disk and point model_path at it. Everything downstream in the generation side depends on this loader existing - start here, verify the pipeline holds at low resolution, then climb.