Flux2Klein HDRI LoRA (CRT AutoDL)

Flux 2 Klein 9B is the distilled image editor of 2026 - the model people reach for to change a photo, and the one that carried the year's best realism recipes once you stack LoRAs on it. Flux2Klein HDRI LoRA (CRT AutoDL) is a very specific add-on for that stack: it auto-downloads the "Klein_9B - HDRI_360_panoramic" LoRA and applies it to your model. If that sounds niche, it is. This node exists for people generating equirectangular HDRI environment maps - the 360° panoramas that light 3D scenes.

What it is

It's the same _FixedLoRALoader pattern as the rest of the AutoDL family: take a model, take a strength_model (default 1.0, range −100 to 100), ensure the LoRA file is present (downloading from the author's own PGCRYPT/Flux2Klein_9B-HDRI repo into models/loras on first use), merge it, output a MODEL. The inputs and outputs are exactly that - two in, one out.

The LoRA itself is trained to make Klein 9B produce 360° HDRI panoramas - equirectangular projections with the wide dynamic range you need for image-based lighting, not just a pretty stitched image. If your workflow is "generate an environment map, drop it in Blender or Unreal as a light probe," this is the node that gets you from prompt to .exr-ready lighting without a detour through CivitAI and a manual file drop.

Is it for you?

Be honest about this one before you install it. If you don't render 3D, you almost certainly don't need an HDRI LoRA - a regular Klein generation won't become a usable light probe, and this LoRA's whole job is making the panorama and the dynamic range correct for compositing. If you do do environment-map work, it's a genuinely nice find, because the source repo is obscure enough that you'd otherwise be digging through Hugging Face search for it. And if you're just here for the realism LoRA stacks the KB talks about (the +1987 Klein realism recipe threads), this isn't that - it's a specialized tool that happens to live in the same downloader family.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search "CRT-Nodes", restart. First run downloads the LoRA to models/loras; subsequent runs load instantly. Uses ComfyUI's core LoRA loader, so no extra dependencies beyond the pack.

Gotchas

The LoRA is trained for the 9B Klein - don't expect it to make sense on Flux 2 Dev or another checkpoint; load_lora_for_models may fail outright if the shapes don't match, which is ComfyUI telling you the LoRA doesn't belong on that model. strength_model at 1.0 is the sane default; if the HDRI look comes out too strong or the panorama seams show, drop it toward 0.5 rather than fighting it with prompts. And remember it only touches the model - you still need the Klein VAE and text encoder from the pack's other AutoDL nodes to actually sample.