SMF·PEBBLELY
Can a prompt replace Pebblely?
AI image generation — AI image and video generation
Exhibit tracking slip
Verdict
Pebblely takes the position control away on purpose: you upload a cutout, choose a look — marble, kitchen counter, beach — and get a grid of variants. That is a batch generation loop with saved prompt themes, and locally it is one sitting on top of ComfyUI. The gap is speed and consistency. Eight variants arrive in seconds hosted and in several minutes locally, and a general checkpoint produces one usable result in eight where a tuned model produces four.
Exhibit A — The prompt
Received on31.07.2026Build a local product-background generator organised around saved themes and variant grids. No canvas, no spatial control.
Input: a product image. Run a local segmentation model (rembg) to produce a cutout with an alpha channel, then refine the matte with erode and feather controls, since the halo left on light backgrounds is what makes a composite obvious.
Themes: a theme is a saved, named bundle — a positive prompt, a negative prompt, a checkpoint, a sampler and step count, and a lighting direction. Ship no themes; the first run walks the user through creating one from a real product, because a theme tuned to your own catalogue beats a generic one. Themes are stored as files so they can be shared.
Generation: for a chosen theme, generate a grid of eight variants at once by varying only the seed. Composite the product cutout onto each generated background at a consistent scale and position derived from the cutout's own bounding box, so every variant frames the product identically and the only thing changing is the scene. Submit to a local ComfyUI instance configured in the environment.
Lighting match: apply a gradient to the composited product matching the theme's declared light direction and strength, and generate a contact shadow from the alpha matte, offset accordingly. Without this the product reads as pasted on, which is the single difference between a usable result and an obvious one.
The grid, which is the interaction: eight results shown at once, click to enlarge, keyboard to keep or discard, and a rerun that regenerates only the discarded slots with new seeds. Keeping a variant records its seed in the theme's history, so a theme that produces good results accumulates known-good seeds worth reusing on other products.
Batch: apply one theme to a folder of products in one run, with a contact sheet of every result grouped by product, so a catalogue of thirty items can be reviewed in one pass.
Reproducibility: write a sidecar JSON per output recording the theme, seed, checkpoint, matte settings and workflow, so any result can be regenerated exactly.
Output: full resolution PNG and JPEG, plus square and vertical marketplace crops.
Out of scope: positioning props or controlling composition, any hosted API, an account system, model fine-tuning, and publishing to a marketplace.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 models tuned for product photography, which is most of the hit rate
- B.2 the generation speed that makes a variant grid worth browsing
- B.3 the curated theme library, kept current with what sells
- B.4 the commerce workflow that pushes results straight to a listing
Prior art
Exhibit C — Why people still pay: frontier models, compute, and data
Because the value is in the hit rate. Nineteen dollars for eight usable backgrounds in ten seconds beats an evening of local generation where two of eight are usable, for anyone whose time has a price.
Questions
Can I bring my Pebblely themes across?
No. Its themes are platform presets against its own models with no export, so you build your own — which takes an evening of experimentation per theme and then keeps working for every product you own.
How many of the eight variants are actually usable?
With a general SDXL checkpoint, expect one to three, against four or five from the paid product. Hit rate is the honest measure in this category, and it is where a tuned model shows its value most clearly. A well-tuned personal theme narrows the gap considerably.
What does it cost to run?
Nothing per image once the models download, but a GPU with 8 GB of VRAM is effectively required — eight SDXL variants on CPU is tens of minutes, which makes the grid interaction pointless.
What is the one thing that does not survive the rebuild?
Speed. A variant grid only works if it arrives while you are still looking at it. Ten seconds means you try five themes; four minutes means you accept the first tolerable result from the first theme, which is a different and worse product.
Related tools
Receipt