LoRA Dataset Studio
Documentation · Installation · Settings
The app scales from "no GPU at all" to a full local training rig — each capability has its own floor, and missing pieces are hidden or guided through Setup.
| Mode / capability | GPU (NVIDIA) | Disk | Notes |
|---|---|---|---|
| API-only (Gemini/ChatGPT/OpenRouter generation, import/scrape, curate, manual captions, export/backup) | none | ~2 GB | Any machine with Python 3.10+ (3.13/3.14 run the core app fine — the 3.10–3.12 window is an ML-extras constraint); Docker image available |
| Auto-captioning & framing (Ollama vision, 8B model) | ~8 GB VRAM | ~7 GB | Runs alongside generation, not concurrently |
| Local generation (Klein 9B KV fp8 via ComfyUI) | ~16 GB VRAM | ~30 GB (model + text encoder + VAE) | Free, local and NSFW-capable; Setup downloads the models. The KV build is up to 2.5× faster on multi-reference edits at the same quality. Available in Docker GPU mode |
| LoRA training — Z-Image / SDXL (ai-toolkit) | 16 GB+ recommended | 10 GB+ free enforced per run | Quantized (qfloat8) + low-VRAM mode |
| LoRA training — Krea 2 (ai-toolkit) | 24 GB VRAM at 1024 px (enforced warning) | ~24 GB base download (Raw), or none if you start from a Krea 2 checkpoint you already have, + 10 GB+ free | Under 24 GB, select Resolution → 768 only in Advanced options |
| LoRA training — FLUX.2 Klein (ai-toolkit) | 4B: 16–24 GB VRAM · 9B: 32–48 GB | base download + 10 GB+ free | Both bases are gated on Hugging Face; the cloud lane is practical for 9B |
| LoRA training — FLUX.1 / Anima (ai-toolkit) | ~24 GB VRAM (both are 12B-class families) | base download + 10 GB+ free | Local only — neither has a cloud lane. FLUX.1 is gated on Hugging Face; Anima's base is public and reads booru tags natively |
| Full-model (dense) training — Krea 2 (experimental, cloud only) | 80 GB VRAM — there is no local lane | 200 GB on the pod, plus a private Hugging Face repo for the ~26 GB master | Not the recommended path: Krea's own advice is a LoRA on Raw applied to Turbo. Needs a separate HF_CLOUD_TOKEN
|
| Face scoring / person masks / watermark inpaint (ML extras) | none (CPU) | ~3 GB (+ CPU torch for LaMa) | Python 3.10–3.12 required for wheels; installable per capability from Setup |
-
OS: Windows 10/11 for the full local stack (
start.bat). Linux/macOS work for API-only + manual venv; GPU Docker depends on host NVIDIA support. -
Python: 3.10–3.12, but not required up front:
start.batfetches a self-contained CPython 3.12 when none is installed. Python 3.13+ can run the core app but not the ML extras. - RAM: 16 GB+ recommended for local training. Unlike VRAM and free disk, this one is a recommendation the app never measures — a run that dies for want of system memory has no guard-rail in front of it.
- Dataset size: a launch is gated on a per-family floor — 12 images for Z-Image, 15 for Krea 2 / FLUX.1 / FLUX.2 Klein, 20 for SDXL, 4 for a slider LoRA — with 20-30 recommended. Below the floor the app asks you to confirm and warns about overfitting rather than refusing outright.
- Reference development rig: RTX 4090 (24 GB); every number above was measured or enforced there.
Core tools are prepared in Setup. Install optional features from Plugins → Store, then use each plugin’s settings and preparation screen. Missing dependencies keep the affected feature unavailable.
| Feature | Requires |
|---|---|
| Nano Banana Pro generation | GEMINI_API_KEY |
ChatGPT / gpt-image-2 generation |
OPENAI_API_KEY, or the separate experimental ChatGPT-subscription connection |
| OpenRouter generation |
OPENROUTER_API_KEY plus an image-capable model slug; OpenRouter billing and the upstream provider's policy still apply |
| Klein generation / improvement | ComfyUI reachable + Klein model stack |
| SeedVR2 upscaling | Reachable ComfyUI, the ComfyUI-SeedVR2_VideoUpscaler node pack and two model files (~3.9 GB). Prepare SeedVR2 installs the nodes, dependencies and models on supported Windows portable layouts; other layouts require upstream node preparation. Re-check readiness after restarting ComfyUI. Optional Comfyui_TTP_Toolset enables overlapping tiling, with always/never controls. See preparation and tiling limits
|
| Krea 2 Edit generation | ComfyUI reachable + comfyui-krea2edit, a Krea 2 base, Identity Edit LoRA, Qwen3-VL encoder and Qwen Image VAE; exact files
|
| Captioning | A local LLM — Ollama or LM Studio — or ai-toolkit (JoyCaption) |
| Dual long + short captions | ai-toolkit + local vision caption derivation; local training only, and unavailable for Krea 2 / Anima |
| Auto-framing / auto head-crop | A local LLM (Ollama or LM Studio) with a vision model |
| Face similarity / auto-triage |
backend/requirements-ml.txt (InsightFace + ONNX Runtime) |
| Character person masks |
backend/requirements-ml.txt (rembg); Concept/Style intentionally disable them |
| Image Bank scoring, crops and semantic tools | The Bank scoring extra provides CLIP and ✨ Score. Each Bank can instead select the optional pinned SigLIP 2 engine from Setup; it builds a separate index, while aesthetic/NSFW/style/medium remain on CLIP. Balanced picks also need Framing. Both ship CPU-only PyTorch on purpose; on a machine that already has a CUDA Python (ai-toolkit's, ComfyUI's) each can be pointed at it instead — checked package by package, never installed into, and separately for ✨ Score and for SigLIP 2. |
| Watermark detection | A local LLM (Ollama or LM Studio) with a vision model, or the dedicated detector (torch + transformers — the bank-scoring extra's environment is reused when present — plus ~0.9 GB of model downloads at first use) |
| Watermark inpainting | LaMa extra from backend/requirements-ml.txt, or ComfyUI + Klein, which erases the found zones and re-renders the whole photo; crop remains model-free |
| Scraping |
backend/requirements-scrape.txt; Pexels also needs PEXELS_API_KEY and explicit authorization. Gallery/URL scanning goes through gallery-dl for any site it recognizes, whatever its bundled extractors cover; an unrecognized site returns "No images found" in the picker (the single item gallery-dl's yt-dlp fallback can still fetch is video-typed, so it never reaches the image list), and a listing of albums returns one cover per album unless Scan full albums is ticked. A scan that was cut short — by the time budget, a result cap, or a source that blocked or rate-limited it — now says so under the results ("this scan stopped before the end of the listing"), instead of presenting a partial list as the whole thing. Web image search needs no key — it queries a metasearch layer over several backends and asks for photos, but the filter is not honored uniformly, so some non-photo results can still come through; results are capped per search rather than guaranteed — a request for the 120 maximum routinely comes back with far fewer — come from third-party sites whose licence is your responsibility, and a few links — mainly stock-photo CDNs that redirect to the actual file — are refused by the hardened fetch that protects every import |
| Video Bank — reading and triaging |
backend/requirements-ml.txt (PyAV). Shot detection additionally needs transnetv2-pytorch (weights bundled, nothing to download), which rides the bank-scoring environment because it pulls torch. The three pieces install and fail apart, and Setup reports them as three separate rows |
| Video Bank — cutting clips into a dataset | An ffmpeg binary: imageio-ffmpeg ships one, or any ffmpeg on PATH. Needed only to promote — without it you can still scan, detect shots, watch and triage a whole bank |
| Video Bank — shot captions and scene search | The Bank scoring extra's environment (torch + transformers ≥ 4.57) plus a Qwen3-VL checkpoint downloaded at first use; the model is a setting, and the same environment serves ✨ Score, SigLIP 2 and the watermark detector |
| Civitai scanning |
backend/requirements-scrape.txt; without CIVITAI_API_KEY the scan runs but returns SFW results only |
| 🌐 Civitai top prompts (Studio/Canvas) | Browsing needs nothing; reading the prompts needs CIVITAI_API_KEY (free account) — the same key Civitai scanning uses |
| 📷 Camera angles | ComfyUI reachable + the Qwen-Image-Edit stack Setup's Camera card downloads (the VAE is shared with Krea 2 Edit) |
| 🔤 Find text (bank & dataset) | The same small CPU OCR package the Video Bank's text pass uses, installed from Setup |
| Local LoRA training: Z-Image / Krea 2 / FLUX.1 / FLUX.2 Klein / Anima / Qwen-Image 2.1 | ai-toolkit; Qwen-Image 2.1 requires a recent toolkit with qwen_image_2 support. No ComfyUI is needed for official Hugging Face bases. Krea 2 can also start from a local checkpoint discovered through ComfyUI's model tree, including a delivered full model. Ordinary fp8 weights are up-cast with a precision warning; packed ComfyUI exports are refused because the trainer cannot load their decompression tables |
| Local SDXL training | ai-toolkit + a base checkpoint discoverable in ComfyUI's model tree |
| Cloud training |
VAST_API_KEY; supported families are shown in the launch UI. Full-model Krea 2 also needs HF_CLOUD_TOKEN with Krea base read and repository write access; fine-grained is recommended, global role=write is accepted with a warning, and read-only is rejected. A finished full model (~26 GB, plus its ~10 GB fp8 twin) is downloaded to your machine and verified before the rented pod is released; a copy of the master is then pushed to a private Hugging Face repository as a backup, and it can be turned off. Either copy can seed a fresh pod when you continue the run: the Hugging Face one is minutes over a datacenter link, the one on your machine costs your upload speed — and the ▶ Continue dialog shows both durations and what each one costs in rented GPU time before you pick. So you need room on the checkpoint drive (checked before anything is rented), and Hugging Face room only for the backup; Settings ▸ Storage lists what is taking that space |
| Quantizing a model to fp8 (Settings ▸ Storage, or a full model's card) | A Python with torch; the interpreter is probed before the button is enabled, so a missing package is a refusal with its pip line, not a crash thirty seconds in. Runs on the CPU, one at a time, so it never takes VRAM from ComfyUI or a training run. Your source file is never modified or overwritten; an already-quantized file, or an adapter, is refused |
| Merging a LoRA into a base checkpoint (produces a full model) | A Python with torch (the same one fp8 quantization uses) and room for a second copy of the base — a 26 GB Krea 2 base takes about two minutes and writes 26 GB. Refused on an already-quantized base: merge into the full-precision file, then quantize. LoRAs must name their modules the way the base names its weights (the ai-toolkit/diffusion-model convention); kohya's flattened lora_unet_… SDXL exports do not, and are refused by name before anything is written. The result is a merged model, not a trained one, and its metadata says so |
| LoRA Canvas browsing, layout, notes and diffs | No external service; generating needs ComfyUI and same-family checkpoints, continuing needs the chosen local/cloud training lane |
| Test Studio | ComfyUI reachable + assets for a supported Studio family |
| Backup/restore and ZIP/folder merge | No external service |
| Hugging Face publishing | Write-enabled HF_TOKEN; repositories are private by default |