Pony LoRA Training Requirements
Pony LoRA training requirements begin with a compatibility decision that generic SDXL checklists often miss. Pony Diffusion V6 XL is based on SDXL and can use SDXL LoRA tooling, but it has its own learned score, source, and rating vocabulary. Pony V7 belongs to an AuraFlow-based lineage. A V6 adapter and a V7 adapter are not interchangeable even when both are described casually as Pony LoRAs.
This page is a readiness checklist for Pony V6 and compatible V6-derived checkpoints. It does not repeat the dedicated Pony workflow guide. Use it to confirm the exact model family, trainer path, GPU and storage plan, 1024-class dataset, tag convention, rights records, checkpoint schedule, and Pony-specific validation prompts before training. The central requirement is alignment: architecture, captions, prompts, and inference checkpoint must all belong to the same documented branch.
Quick Answer
For a Pony V6 LoRA, you need a known Pony V6 or compatible SDXL-derived checkpoint, a maintained SDXL LoRA trainer, enough memory for the complete SDXL configuration, a clean bucketed dataset, a stable trigger, and captions designed for Pony's vocabulary.
Decide how you will use quality score tags, one accurate source tag where appropriate, and rating vocabulary. Save intermediate checkpoints and validate with Pony prompts both with and without those conditioning tags. Do not use this checklist for Pony V7, which is based on AuraFlow and requires explicit trainer and adapter support for that architecture.
Requirement 1: Prove the Pony Branch
Record the exact checkpoint repository, filename, revision or hash, model card, base model, and license.
Pony Diffusion V6 XL identifies as an SDXL fine-tune. It uses SDXL-compatible loaders and training scripts. Pony V7's documented base is AuraFlow, so V6 scripts, adapters, and assumptions do not transfer automatically.
A checkpoint name containing "Pony" is not enough. Community derivatives can be merges, remixes, or models that borrow a prompting style without preserving the same lineage.
Confirm:
- V6 or SDXL ancestry is stated in the model card;
- the intended trainer supports SDXL LoRA training;
- the inference checkpoint belongs to the same V6 family;
- the model's recommended CLIP, VAE, and prompt behavior are documented;
- the license permits the planned training and distribution.
Requirement 2: A Pony-Aware Prompt Vocabulary
The V6 model card documents a quality chain beginning with score_9 and descending _up tags, source selectors such as source_pony, source_furry, source_cartoon, and source_anime, plus rating tags.
This vocabulary affects both dataset design and validation. Before training, write a convention that answers:
- Will quality tags appear in captions, only in validation prompts, or both?
- Which source tag accurately describes each image?
- Will one rating tag be included where relevant?
- Will content use booru-style tags, natural language, or a controlled hybrid?
- Where will the trigger appear?
The model card says Pony V6 was trained with both tags and natural-language descriptions, so either format can work. Consistency and accuracy matter more than maximizing tag count.
Requirement 3: Deliberate Score Tag Handling
The documented V6 generation prompt uses a score chain such as score_9, score_8_up, score_7_up, score_6_up. Those tokens are conditioning vocabulary, not objective quality certification.
Repeating the same score chain in every training caption can bind it to the trigger. Omitting score tags may give the adapter less direct exposure to the checkpoint's learned quality vocabulary. There is no universal choice for every dataset.
Pick a policy, record it, and create validation prompts that compare:
- the full chosen score chain;
- a shorter score chain;
- no score tags;
- the trigger with otherwise identical content.
If the LoRA only works under one score string, document that dependency.
Requirement 4: Accurate Source and Rating Tags
Use source tags only when they match the image. source_anime and source_furry are not generic quality modifiers. Wrong source labels add contradictory conditioning.
Rating vocabulary must also be accurate and lawful. A rating tag describes learned conditioning. It does not grant permission to use an image or create prohibited content.
For safe projects, keep the dataset and validation suite within the intended safe scope and use the model's documented safe conditioning where appropriate. Do not train non-consensual intimate imagery, exploitative material, or deceptive identity content.
Requirement 5: SDXL-Capable Compute
Pony V6 follows SDXL memory behavior. The exact requirement depends on:
- 1024-class resolution and bucket dimensions;
- physical batch and accumulation;
- mixed precision;
- rank and alpha;
- optimizer state;
- UNet-only or text-encoder training;
- gradient checkpointing;
- latent and text-output caches;
- validation dimensions.
NVIDIA CUDA has the broadest support in current sd-scripts and related interfaces. Other backends depend on the trainer. Verify GPU precision support, driver, runtime, system RAM, and free SSD space.
Do not claim a fixed VRAM minimum from one preset. Complete a smoke test that includes cache creation, backward passes, checkpoint saving, and Pony validation sampling.
Requirement 6: A Current SDXL Trainer
Kohya-ss sd-scripts uses sdxl_train_network.py for SDXL-family LoRAs. Diffusers maintains SDXL LoRA examples. GUI wrappers may expose those settings, but the backend release still matters.
Save:
- trainer name and commit or release;
- Python and PyTorch versions;
- driver and CUDA version;
- architecture selection;
- adapter module and export format;
- all memory controls;
- sample and checkpoint settings.
A generic SD 1.5 trainer path is not sufficient. A FLUX preset is not sufficient. Pony V6 needs the SDXL path.
Requirement 7: A Conservative Component Strategy
SDXL has two text encoders. Current sd-scripts guidance supports a UNet-only first LoRA, which is a sensible Pony baseline.
Text-encoder training should have a defined purpose, such as testing a trigger or specialized vocabulary that the frozen conditioning fails to represent. It raises memory use and complexity.
If text-encoder outputs are cached in sd-scripts, caption shuffling and related augmentations are restricted, and text-encoder LoRA training is unavailable in that cached run. Your tag strategy must agree with those settings.
Do not enable caption shuffling in a planning note and then cache fixed text outputs in the actual run.
Requirement 8: A 1024-Class Pony Dataset
Use SDXL-style resolution and aspect-ratio buckets supported by the trainer. Inspect processed crops. Pony's stylized subjects often include ears, tails, wings, horns, hands, feet, clothing details, and unusual silhouettes that automatic crops can remove.
Require:
- consistent target identity or style;
- varied pose, angle, expression, framing, outfit, and background;
- clear anatomy and geometry;
- duplicate and near-duplicate removal;
- no unwanted watermarks, signatures, text, or borders;
- sufficient real detail at training scale;
- documented image rights and consent;
- an immutable source backup.
For a style LoRA, vary species, characters, scenes, and compositions. For a character LoRA, preserve identity while changing incidental details.
Requirement 9: Audited Pony Captions
Every final image needs the caption format expected by the trainer. Use a consistent order that makes review easy, for example:
score tags, source tag, rating tag, trigger, subject count, identity features, clothing, pose, framing, scene
This order is an audit convention, not a secret training formula.
Auto-taggers are useful drafts, but stylized anatomy creates false positives. Review:
- character count and species;
- hair, fur, skin, eye, and marking colors;
- ears, tails, wings, horns, and limbs;
- clothing layers and accessories;
- pose, expression, and camera framing;
- source and rating tags;
- trigger spelling;
- synonyms and underscore formatting.
Remove artist names unless you have a clear lawful reason and rights to use them. Remove tags that are not visible.
Requirement 10: A Versioned Dataset and Exposure Plan
Freeze a manifest containing filename, hash, source, rights status, dimensions, caption review status, and inclusion decision.
Define repeats, epochs, batch, accumulation, and total steps together. A small dataset with high repeats can overfit quickly. Save several intermediate adapters rather than selecting the final step automatically.
Do not change tags halfway through a run you intend to compare. Create a new dataset revision and run ID.
Requirement 11: Pony-Specific Validation
A generic SDXL sample prompt does not prove that a Pony adapter works. Build a fixed validation grid using the target checkpoint's documented vocabulary.
Include:
- the trigger with the planned quality chain;
- the trigger without score tags;
- each relevant source tag;
rating_safefor safe output checks where applicable;- new outfits, backgrounds, expressions, and poses;
- close, medium, and wide framing;
- several adapter strengths;
- a prompt without the trigger;
- the untouched base checkpoint;
- one claimed compatible V6 derivative.
Keep seed, sampler, steps, guidance, resolution, and prompt text fixed across checkpoints. Judge identity or style fidelity, anatomy, prompt control, source-tag response, and unwanted composition leakage separately.
Requirement 12: Explicit Compatibility Boundaries
The final model card should say Pony V6 / SDXL and identify the exact training base. List the trigger, expected score and source tags, recommended strength, tested checkpoints, trainer version, resolution, known failures, and license.
State that Pony V7 or AuraFlow is not supported unless a separately trained and tested adapter exists. Do not imply portability based on the shared Pony name.
If the adapter technically loads on a non-Pony SDXL checkpoint but performs poorly, do not call that checkpoint compatible.
Pony Preflight Test
Run a short end-to-end job:
- Generate a baseline image from the exact Pony checkpoint.
- Parse every image and caption.
- Inspect bucket counts and crops.
- Build intended caches.
- Train for several steps with finite loss.
- Save an adapter.
- Render the Pony validation prompt.
- Reload the adapter in the intended inference workflow.
Check GPU and system memory, sample integrity, trigger response, tag response, and checkpoint compatibility. A trainer process reaching the final line is not enough.
Common Missing Requirements
The checkpoint is vaguely labeled Pony: ancestry and license were never verified.
The dataset has generic anime tags: score, source, rating, and checkpoint vocabulary were not planned.
Every image has the same quality chain: the trigger may become coupled to that conditioning.
Caption shuffling does nothing: cached text outputs fixed the conditioning earlier.
The adapter works only on one merge: training base compatibility was too narrow and undocumented.
V7 output fails: a V6 SDXL adapter was applied to an AuraFlow branch.
The character keeps one outfit: clothing lacked variation or accurate tags.
Bottom Line
Pony V6 readiness means more than selecting SDXL in a trainer. Verify the branch, use Pony-aware captions, make score and source tag choices explicit, preserve 1024-class coverage, and test with Pony prompts on the actual inference checkpoint.
Keep V6 and V7 separate. Clear architecture and vocabulary boundaries prevent wasted runs and misleading compatibility claims.