FLUX.2 Klein Base 4B LoRA · Edit

Here's a node with a strange price tag: Klein Base 4B LoRA · Edit costs 3.6 credits ($0.036) an image, which is more than the distilled Klein 4B LoRA Edit at 3.06 - for a checkpoint that's slower and, by common agreement, worse for producing pictures. That's not a pricing bug. It's the cost of an undistilled 50-step trajectory on a model built to be trained on rather than used. So who's it for? Basically: people testing a fine-tune.

The Base distinction, quickly

Klein shipped as four checkpoints on two axes - size (4B/9B) and distillation (distilled/base). The distilled pair run four steps and are what everybody generates with. The Base pair run a full undenoised trajectory, which is why BFL's own guidance is that Base is the foundation for training, not a generation model. The pack's summary for this lane is "undistilled Klein 4B that loads your own LoRAs" - accurate, and note what it doesn't claim.

On edits, Base is interesting for one reason: your LoRA was probably trained on it. If you fine-tuned a 4B adapter for a specific product, material, or grading look, this endpoint is the closest thing to the conditions it learned under. You can then move the same file to the distilled nodes and see how much of the behaviour survives the distillation. That comparison is the whole point, and at three and a half cents a look it's cheap.

Loading the weights

loras is a POPNOW_LORAS socket fed by the pack's PopNow LoRA node. Its source accepts a .safetensors in ComfyUI/models/loras (uploaded once, reused about 12 hours), a Hugging Face repo id or page, a Civitai page (imported server-side - CIVITAI_API_KEY or HF_TOKEN in the environment for gated files), or a direct https:// link. scale is the strength dial. Chain nodes to stack up to three.

The tooltip here accepts files trained on Base 4B or 4B - same size, either build. What it won't accept is a 9B adapter (wrong size) or anything from the FLUX.1 generation (wrong architecture entirely; Flux 2 reads prompts with a Qwen3 encoder, not CLIP+T5).

Editing fields

image_urls is Autogrow - one socket to start, another each time you connect, three maximum. The service refers to them positionally as image 1, image 2, image 3, so write instructions that use those numbers. Keep one image per socket unless you mean it: a batch counts as one reference per image, and five references into a three-slot model earns you a PopNow 400.

prompt is 5,000 characters of plain instruction, trigger word first, change plus invariants. aspect_ratio starts at auto and keeps your first source's shape. resolution is 1k or 2k; editing endpoints bill flat, so the larger tier isn't charging extra at the moment. seed sits on fixed - the reason a re-queue looks like nothing happened - so flip control after generate to randomize when you're comparing runs.

Outputs

images → Save Image / preview / your post chain. task id → the PopNow job handle for the API console and support.

Install

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

Manager → Custom Nodes Manager → PopNow → Install does the same, with a restart either way. Key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY in the environment), then restart once more so the catalog loads from your account. The pack has no dependencies of its own, but it needs a ComfyUI new enough for the V3 node API it's written against.

Don't mistake this for an upgrade

If you're choosing blindly between the four Klein editing nodes, pick by price: 4B Edit at 1.89 for drafts, 9B Edit at 2.44 for real work, 4B/9B LoRA Edit when you have weights. Come here for the specific job of validating a fine-tune on its native base, and expect the output to look a little less "finished" than the distilled nodes - that's the checkpoint being honest, not your LoRA failing.