FLUX.2 Klein Base 9B LoRA · Edit

Klein Base 9B LoRA · Edit is the most expensive Klein node in this pack at 4.68 credits ($0.0468) an image - nearly double the plain Klein 9B Edit, and more than the distilled LoRA edit at 3.78. Every instinct says skip it. Occasionally you shouldn't, and knowing which occasions is the whole point.

What "Base" buys you on an edit

Klein's undistilled checkpoints are the training foundations: 50 steps instead of 4, full trajectory instead of a distilled shortcut, and per BFL's own guidance not the thing to generate with. What they do better is preserve variation - distilled models collapse toward one confident interpretation, undistilled ones keep more of the distribution.

That matters for editing in one concrete scenario. The overwhelming majority of published Klein LoRAs were trained on FLUX.2-klein-base-9B and then applied to the distilled 9B for speed. If you're debugging a LoRA that behaves oddly on edits - over-triggering, bleeding into the background, losing a face it should hold - reproducing the failure on the native base tells you whether the problem is the adapter or the distilled checkpoint it's riding on. When you run an edit here, you're running it under the conditions the adapter was trained in, and that's the diagnostic.

The second, smaller case: a LoRA trained for editing - a consistency adapter, a delighting/relighting LoRA, a texture unwrap - where the deliverable is a faithful transformation rather than a pretty picture. Undistilled output is a slightly less opinionated interpretation of your instruction, which for that class of job is what you want.

If neither of those sentences describes your afternoon, use Klein 9B LoRA · Edit at 3.78 and be happy.

The plumbing, which is the same as everywhere else in this pack

loras takes a POPNOW_LORAS value from the PopNow LoRA node - that's the node with a source field and a scale slider, and its output is what plugs in here. source resolves a .safetensors from ComfyUI/models/loras (uploaded once, ~12h cache), a Hugging Face repo id or page, a Civitai model page (PopNow imports it server-side; CIVITAI_API_KEY / HF_TOKEN from the environment for gated files), or any direct https:// weight link. Chain PopNow LoRA nodes to stack up to three.

The tooltip's exact wording is worth following: files trained for FLUX.2 [klein] 9B, the Base or the distilled build. Within 9B, either works. A 4B adapter will not, and a FLUX.1 LoRA is a different architecture entirely.

Fields

image_urls - Autogrow, up to three sockets, addressed in the prompt as image 1, image 2, image 3. Each socket can carry a batch and every image in it becomes its own reference, so one image per socket is the safe habit. prompt - 5,000 characters, trigger phrase first, then what changes and what must stay; no weighting syntax, the encoder ignores it. aspect_ratio - auto keeps your first source's shape. resolution - 1k or 2k, billed flat on the edit endpoints, so the larger tier is currently free. seed - fixed until you deliberately flip control after generate.

Outputs

images is a standard IMAGE batch; task id is the string for the API console and support.

Install

Manager → Custom Nodes Manager → PopNow → Install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui
# restart ComfyUI

Then a key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY in the environment on a server) and one more restart, since the node list is read from your account's catalog at startup. No pip dependencies at all - but the pack is built on ComfyUI's V3 node API, so an old install will fail before it ever offers you a node.

A last word on the price

Five cents per edit only looks expensive until you compare it to the alternative, which is running a 9B undistilled model plus a Qwen3-8B text encoder locally at 20-plus GB. That said, exploration on this node is a waste of credits: do your framing and wording on the 2.44 node, and come here to answer a question - does my adapter work on its own foundation - then go back. Set Settings > PopNow > Max credits per run early; it's the only thing standing between you and a loop that quietly empties a wallet.