SMF·PLAYGROUND-AI
Can a prompt replace Playground AI?
AI image generation — AI image and video generation
Exhibit tracking slip
Verdict
The generation itself runs locally on open weights, and the workbench around it — a searchable history where every image keeps its full parameters and any past generation can be forked with one setting changed — is genuinely a sitting's work and genuinely useful. What you cannot rebuild is the reason people open Playground: models tuned for a house style, iteration fast enough to explore, and a feed of other people's prompts to learn from.
Exhibit A — The prompt
Received on31.07.2026Build a local image-generation workbench whose distinguishing feature is a searchable, forkable history.
Generation: submit jobs to a local ComfyUI instance configured in the environment. Expose prompt, negative prompt, seed, dimensions, steps, sampler, CFG scale, checkpoint, and LoRA selection with weights. Support image-to-image and inpainting with a mask drawn in the interface.
History, which is the point of this build. Every generation writes a record holding the output image, every parameter above, the workflow JSON, the checkpoint file hash, generation duration, and a timestamp. Nothing is ever silently discarded — a generation you rejected is exactly the one you will want to find again next week.
Search over that history: full-text over prompts and negative prompts, plus filters on checkpoint, LoRA, seed, dimensions, and date. Search results as a contact sheet. This is what turns a folder of PNGs into a working archive.
Forking: from any past generation, open a new job pre-filled with its exact parameters, change one thing, and run. Record the parent id so a lineage is preserved, and show any generation's ancestry as a chain — this image came from that one with the seed changed, which came from that one with a different LoRA. Being able to walk backwards through a chain of small changes is the single most useful thing this build offers over a plain queue.
Comparison: select any two generations and view them side by side at 1:1 with their parameter differences highlighted, so "what actually changed" is answerable rather than remembered.
Collections: tag generations, mark favourites, and export a collection as a folder of images with a JSON manifest of parameters, so a set can be reproduced on another machine.
Queue: a visible job queue with per-job progress, estimated VRAM, cancellation, and a clear failure when a checkpoint or LoRA referenced by a forked job is missing locally.
Storage: SQLite for records, images on disk in a dated tree. Document the schema, and provide a rebuild command that reconstructs the index from the sidecar files on disk, so the database is never the only copy of anything.
Out of scope: any hosted API or model, a community feed or sharing, model training, and moderation of generated content.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the tuned proprietary models and their consistent look
- B.2 hosted GPUs, so iteration is minutes rather than seconds
- B.3 the community feed of prompts and settings to learn from
- B.4 the canvas with generative fill across a large composition
Prior art
Exhibit C — Why people still pay: frontier models, compute, and data
Because image generation is a search problem and speed is the search. Fifteen dollars for enough compute to try two hundred variations beats a local card that manages twenty in the same hour.
Questions
Can I bring my Playground history across?
Generated images download, but the parameters behind them are not exported in a structured form, so the history arrives as pictures without their settings. Since the settings are the whole value of a history, in practice you start fresh.
How does local generation compare on quality?
For general images, open SDXL-class checkpoints are close and sometimes better with the right LoRA. Where Playground pulls ahead is consistency — its models were tuned for a coherent look, so a series of images belongs together in a way that a general checkpoint with a long prompt struggles to match.
What does it cost to run?
Nothing per image once models are downloaded, but a GPU with at least 8 GB of VRAM is effectively required. On CPU the loop still works and takes minutes per image, which defeats the exploratory workflow this tool exists for.
What is the one thing that does not survive the rebuild?
Iteration speed. Generation is a search over a space you cannot see, and hosted compute lets you sample it densely. A local card samples it sparsely, which means you settle on a worse result and never find out what you missed.
Related tools
Receipt