Selling Fine-Tunes

Who Sells on Thumper

The marketplace is built for creators who have domain expertise and want to monetize it through trained models:

• Artists — style LoRAs trained on original artwork, texture packs, lighting presets
• Writers — LLM adapters for specific genres, tones, or narrative styles
• Voice actors — voice clone models trained on licensed recordings
• Domain experts — fine-tuned models for legal, medical, financial, or technical domains
• Prompt engineers — textual inversions, embedding bundles, and workflow packages

What You Can Sell

LoRA adapters, merged GGUF models, voice clone checkpoints, textual inversions, and bundles that combine multiple assets with workflows and prompts. All model types that Thumper’s inference engines support are eligible.

Revenue Model

Sellers set their own prices in whole US dollars. Thumper keeps a 15% fee on each sale and handles payment, tax and receipts. Each sale is held for 30 days, the length of the refund window, then paid out on the next Monday to your payout account ($25 minimum), once your identity and tax form are verified. Free models are allowed and encouraged — they bring people to your paid listings.

Bundle Strategy

A bundle combines a trained model with everything a buyer needs to get results immediately: sample prompts, recommended settings, a ComfyUI workflow, and preview images showing what the model can do. Buyers are paying for time savings, not just weights.

The most successful sellers offer a free standalone LoRA alongside a paid bundle. The free version drives discovery and reviews; the bundle captures buyers who want the full experience.

What to Include

• Adapter or model file — the trained weights in the appropriate format (safetensors, GGUF, pth)
• Preview images or samples — at least 5 high-quality outputs showing the model’s capabilities
• Trigger words and prompts — the activation keywords and example prompts that produce the best results
• Base model information — which base model and version the adapter was trained on
• Recommended settings — weight strength, sampler, steps, CFG scale for image models; temperature and top_p for LLMs
• Workflow file (bundles) — a ComfyUI workflow JSON or configuration that uses the model correctly

Quality Checklist

Before submitting, verify your model meets these quality standards:

• Test with at least 10 different prompts covering the model’s intended use cases
• For image LoRAs: test at weight strengths 0.5, 0.75, and 1.0 to document the usable range
• For image LoRAs: test with at least 3 different samplers (Euler, DPM++ 2M, DDIM)
• For LLM adapters: test with the base model’s default settings and your recommended settings
• Verify the model works with the base model version specified in your manifest
• Check that trigger words are unique and do not conflict with common vocabulary

Every model listing requires a manifest file that describes the model, its lineage, configuration, and content metadata. The manifest follows Thumper’s model manifest v2 schema.

# my-style-lora.model.yaml
schema_version: 2
id: my-studio/watercolor-dreams-v2
name: Watercolor Dreams v2
description: >
  Soft watercolor painting style with bleeding edges,
  visible paper texture, and muted color palettes.
  Trained on 200 original watercolor paintings.
kind: style-lora
version: "2.0"
 
lineage:
  base_model: stabilityai/stable-diffusion-xl-base-1.0
  base_model_version: "1.0"
  training_method: lora
  training_tool: kohya_ss
  training_steps: 3000
  training_resolution: 1024x1024
 
adapter_config:
  rank: 32
  alpha: 64
  target_modules: ["to_q", "to_k", "to_v", "to_out.0"]
 
trigger_words:
  - wcdreams
  - watercolor_dreams_style
 
weight_range:
  min: 0.4
  recommended: 0.75
  max: 1.0
 
stacking_config:
  stackable: true
  max_combined_weight: 1.5
  recommended_combined_weight: 1.0
 
original_content:
  declaration: true
  source: "Original watercolor paintings by Studio Name"
  license_holder: "Studio Name LLC"
 
files:
  - filename: watercolor-dreams-v2.safetensors
    sha256: a1b2c3d4e5f6...
    size_bytes: 151126016
 
tags:
  - watercolor
  - painting
  - style
  - sdxl

The manifest is validated automatically during submission. Missing required fields, invalid references, or schema violations are flagged before your listing goes to review.

Every paid listing is a one-time purchase in whole US dollars; the lowest price is $1.00. You choose the license buyers accept (for example your own license, or the base model's license where it allows resale), and the product page names it. Buyers keep what they bought, and your terms can include later releases while the listing stays listed. There are no subscriptions, time-limited or per-seat licenses.

Pricing Guidance

The marketplace is young, so pricing norms are still forming. Here is what we observe from early sellers:

• Simple style LoRAs with minimal documentation sell at $3–$8
• High-quality LoRAs with previews, prompts, and weight guidance sell at $8–$15
• Bundles (LoRA + workflow + prompts + previews) command $15–$50
• LLM adapters for specific domains (legal, medical, coding) sell at $20–$50
• Voice clone models with licensing documentation sell at $25–$75

The submission process ensures quality and safety for all marketplace listings. Follow these steps to submit your model.

1. Create your model manifest YAML file following the schema above
2. Run the manifest validator locally: thumper manifest validate my-model.model.yaml
3. Prepare your preview assets: at least 5 sample outputs, a cover image (1200x630), and a description
4. Choose New › App, model or pack in Thumper and pick Model or Pack
5. Fill in the price in whole dollars, the license and the category
6. Sign and submit — automated checks run first (file integrity, manifest validation, file format)
7. A Thumper reviewer checks content policy, quality and licensing. Your listing shows when you submitted it and its status; we don't promise a review time
8. Once review passes, your listing goes live in the catalog and appears in search results

Automated Checks

Automated checks run within minutes of submission and verify:

• Manifest schema validity and required fields
• File integrity (SHA-256 matches manifest declaration)
• File format validation (correct safetensors/GGUF/pth headers)
• File-format sanity checks
• Preview image dimensions and format requirements

Manual Review

A reviewer checks content policy compliance, quality and licensing. You get an email when review passes or if something needs a fix, and the listing always shows its status. Every new version is reviewed again before buyers get it.

Your earnings page shows each sale, the 15% fee, what is on hold and what has been paid out. One fee applies to every seller.

Growth Tips

• Respond to reviews — reply to feedback and fix what people report
• Update for new base models — when a new SDXL or Qwen version drops, retrain and release a new version
• Cross-promote — link related models in your descriptions (your style LoRA + your character LoRA)
• Create tutorials — a video or blog post showing your model in action helps people decide
• Group complementary models — a style LoRA + matching textual inversions as one model pack