Load AsymFlow Model

Load AsymFlow Model is the entry point for the AsymFlow half of this pack: it takes the FLUX.2 klein Base 9B diffusion model and slaps a ~700MB AsymFlow adapter on top so the model generates pixels instead of latents. This is the node behind the "pixel-space Flux" posts that got people genuinely excited in mid-2026 - a rank-asymmetric flow formulation that restricts noise prediction to a low-rank subspace while predicting data in full dimension, which is what lets you skip the VAE and produce images that look sharper and more "real" than the latent base model.

You reach for this node when you want AsymFLUX.2 klein 9B output: raw, diverse, photorealistic-in-a-good-way generation. Note what it's not - there's no editing support yet (the author is clear about that), and it's not a few-step model. This is a full-quality recipe; the pack pairs it with a ~32–38 step custom schedule, not the 4-step pi-Flow trick.

What it loads

Three things, two of which you choose:

• model_name - the base diffusion model, from your models/diffusion_models folder. The workflow uses flux-2-klein-base-9b-fp8.safetensors.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. The fp8 options load the weights in 8-bit; on a 12–24GB card, fp8 is the pragmatic choice and near-lossless in practice.
• adapter_name (optional) - the AsymFlow adapter from models/loras. Three variants ship: the base asymflux2_klein_9b (rawest, most diverse, most sensitive to prompting) and two SFT variants finetuned on Z-Image Turbo or FLUX.2 klein distilled data (more stable, better anatomy and text, less diversity).
• adapter_strength - default 1.0, range −100 to 100. The author's tip: at high resolutions, push it to 1.2–1.4 to suppress noise.

Output is a single MODEL - which, because it's pixel-space, needs the Oklab Color Encoder standing in as its VAE, plus Clamp Denoised and Pixel Preview to keep it sane in the graph. The shipped asymflux2_klein_9b.png workflow wires all of that up for you.

Install and files

Clone the pack, then download the three model pieces the docs spell out:

cd ComfyUI/custom_nodes
git clone https://github.com/Lakonik/ComfyUI-piFlow

• Base model → models/diffusion_models/flux-2-klein-base-9b-fp8.safetensors (from black-forest-labs on HF)
• Adapter → models/loras/asymflux2_klein_9b.safetensors (Lakonik/AsymFLUX.2-klein-9B)
• Text encoder → models/text_encoders/qwen_3_8b_fp8mixed.safetensors, loaded via a CLIPLoader set to type flux2

Requires ComfyUI 0.17.0+ (current releases pin newer). The README also insists you uninstall any non-official AsymFlow extensions first - community clones exist and collide with this pack's loader internals.

Common issues

The base adapter is famously picky about prompt wording and sampling settings; your standard prompts may come out weird. Keep prompts short with styling cues, and if output is too chaotic, swap to an SFT variant. LoRAs trained on vanilla Klein 9B can still work but mostly at lower strength. If you want GGUF base models, install ComfyUI-GGUF first and the separate Load AsymFlow Model (GGUF) node appears automatically.