FLUX.2 Max · Text to Image

Ten and a half cents a picture. That's six Klein 9B generations, or a whole minute of a cheap video model. FLUX.2 Max is the top of Black Forest Labs' hosted stack - "photoreal light and exact brand colors," in the pack's words - and it is priced like a tool for people who are being paid for the output.

Not sure it's for you? It probably isn't. But the case for it is real, so let's be precise about it.

What Max is, and isn't

Max is API-only. There are no Max weights anywhere; nobody is quantizing it for a 3090 because nobody can. What you're buying is BFL's best-served endpoint: the tier where they don't have to compromise on how many steps the model actually gets, and where colour reproduction is good enough to match a client's hex codes instead of approximating them. The community's reactions to it are the ones you'd expect from a premium API - test threads about pushes to 8K through custom tiling pipelines, and a lot of people asking whether they can fine-tune or LoRA it. (They can't, not through this node and not through BFL. The LoRA lanes in this pack are the ones whose names literally end in LoRA.)

Where does it fit next to the siblings? Dev is the distilled open model at about 1.7 credits. Pro is the undownloadable base at about 4.5. Max is above both at roughly 10.5 at 1k and about 13 at 2k. If you want my blunt read: the jump from Dev to Pro buys you more per cent than the jump from Pro to Max, unless you're doing product work where the exact colour or the exact light rig is the deliverable.

The mechanism, in one paragraph

Queue the graph and the node serialises its fields, asks PopNow for a quote, submits the task, polls until it's done, downloads the image and hands it into your graph as a standard IMAGE. Nothing is sampled locally; your GPU's only job is decoding a PNG. Failed runs are refunded, cancelling stops you waiting but not the job - it keeps running server-side and gets billed if it succeeds, with the result sitting in the API console.

The four inputs

prompt is the entire creative decision, and the field is generous: 5,000 characters. Flux 2's text encoder is a Qwen3 LLM reading an instruction, so the winning style is prose with explicit spatial relations - a matte black kettle on a concrete counter, hard side light from the left, no props - not tag lists and not prompt weighting, which does nothing here.

aspect_ratio gives nine ratios from 1:1 to 21:9 in both orientations. resolution is 1k or 2k, 2,048px on the long side; on Max that step up is roughly a quarter more credits, and the node badge shows it before you commit.

seed begins at fixed by design. The same graph queued twice reuses the first result rather than billing you again, which matters much more at this price. Move control after generate to randomize only when you've actually decided to buy a new roll.

Outputs

images → Save Image / Preview Image / your post chain. task id → a text node if you want the receipt; it's the identifier support will ask for.

Install

Manager → Custom Nodes Manager → search PopNow → Install → restart. Or, from a terminal:

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, restart again so the catalog is read from your account. There is genuinely nothing else: the pack's dependency list is empty and it rides on ComfyUI's own packages. It does require a recent ComfyUI, though - these are V3 API nodes (comfy_api.latest), and an older build will fail to import the pack rather than politely hiding the node.

Don't burn the wallet the first afternoon

Go into Settings > PopNow > Max credits per run and put a number in before you queue anything on Max. A ceiling of, say, 50 credits covers a handful of images and stops a stray 2k batch loop from eating the account before you notice. Every run is quoted first, so the ceiling actually fires before credits are held.

Then the discipline that saves the most money: compose on Klein 4B, which is a tenth the cost, and move a finished prompt up the ladder one tier at a time. Max is a finishing tool. Treating it as a sketchpad is a hobby for people who enjoy reading invoices.