Eric Qwen-Image Multi-Stage Generate

Eric Qwen-Image Multi-Stage Generate is the DIY version of the pack's progressive-generation idea: up to three stages, each one a generate → upscale → re-denoise cycle, with independent control over steps, CFG, and denoise at every stage. It's the node between the plain Generate (single pass, no upscaling) and UltraGen (the tuned, best-practices flagship). If you want to understand what progressive upscaling does before you trust UltraGen's presets, this is the teaching node.

How the stages fit together

• Stage 1 drafts at low resolution from pure noise. s1_mp (0.5 default), s1_steps (15), s1_cfg (8.0). The idea, straight from the KB's upscaling playbook: lock the composition at low resolution where the model is most reliable, then refine.
• Stage 2 upscales the Stage 1 latents (upscale_to_stage2 = area multiplier, 2.0 default) and re-denoises. s2_steps (20), s2_cfg (4.0), s2_denoise (1.0 - full re-denoise preserves little of Stage 1; lower to keep more).
• Stage 3 does the same again (upscale_to_stage3 = 2.0, s3_steps 15, s3_cfg 2.0, s3_denoise 1.0).

The two switches that shape the whole run: upscale_to_stage2 = 0 → output Stage 1 only (single-stage); upscale_to_stage3 = 0 → two-stage output. That makes this one node a whole family: set both to 0 and it's basically a slow Generate; leave both default and you get 0.5 MP → 2 MP → 4 MP-ish.

The inputs that matter

• prompt (required) - write sentences, not tags; the text encoder is Qwen2.5-VL.
• aspect_ratio - applied at every stage, so all stages share the shape. Default 1:1 Square.
• seed - 0 = random.
• s1_mp - the single most important dial. Too high and Stage 1 costs the whole budget; too low and Stage 2 has to invent too much detail from nothing. 0.5 is the sweet spot the pack defaults to.
• s1_cfg vs s2_cfg vs s3_cfg - notice the pattern baked into the defaults: high CFG (8) at the draft to lock composition, medium (4) at refinement, low (2) at polish. That staircase is the whole philosophy; you can fight it, but the defaults already encode it.
• s2_denoise / s3_denoise - your control over how much each stage re-renders vs. preserves. 1.0 = full re-denoise (ignores prior stage structure); lower values refine while keeping composition.

Output is a single image.

The mechanism note worth knowing

Between stages, latents are upscaled with bislerp and re-noised to match the per-stage denoise before re-sampling. That's the difference between this and naive "upscale then one more pass" workflows - the re-noising is calibrated to the upscale factor so Stage 2 isn't just smearing Stage 1. UltraGen (and the optional 2× upscale VAE) refines this further, but Multi-Stage is the transparent version of the same idea.

Where it fits

This is the node to reach for when you want multi-megapixel output but want to tune it yourself rather than take UltraGen's opinionated defaults. It's also lighter than UltraGen in spirit - no seed modes, no sigma schedules, no max_sequence_length knob - which means fewer surprises when you're learning. The cost: you're responsible for the balance. If you want the author's tuned defaults and the extended prompt-token window, use UltraGen; if you want to understand and control every stage, stay here.

One caution from the community side of this pack: multi-stage generation on a 20B model is slow - people with 5090s have reported hour-long UltraGen runs, and Multi-Stage at three stages isn't meaningfully cheaper. Keep s1_mp low and stages at 2 until your prompt is solid. Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/, with the Qwen-Image pipeline loaded upstream.