Publish LoRAs on Lookhouse¶

This guide explains how to publish LoRAs on Lookhouse. It covers the three core actions — listing existing LoRAs, building new variants on the platform engine, and licensing them — along with how rank maps to a catalog role, how pricing and payouts work, and answers to common questions.

Overview¶

You can list LoRAs you have already trained without retraining them. Listing a LoRA on Lookhouse does not remove it from other platforms, and the platform does not train on your weights. You retain ownership of your LoRAs and can remove them at any time.

Key terms:

• Revenue split: You keep 85% of each paid sale; the platform takes 15%. There is no listing fee.
• Pricing: You set the price for each LoRA. Free listings are also supported.
• Payouts: Processed weekly via Stripe, with no minimum balance.
• Setup time: Listing an already-trained LoRA typically takes about 15 minutes.

Prerequisites¶

To list a LoRA, you need:

• A trained LoRA file.
• A supported base model: SDXL, Flux, SD 1.5, Pony, or a custom checkpoint.
• A creator console account.

To build new variants on the platform engine, you also need a training dataset (images, captions, tags, or reference frames). No local GPU is required.

List an existing LoRA¶

A trained LoRA is ready to publish as-is. There is no dataset to re-upload and no retraining step.

1. Add the file. Upload it through the creator console.
2. Set the metadata. You control the filename, trigger words, sample images, and license per LoRA.
3. Publish.

Any rank is supported.

One subject, multiple base models¶

A single listing can hold the same subject trained against several base models. Upload each base-model version to the same listing rather than creating a separate listing per base model. At purchase, the buyer selects the base model they intend to use, and receives the version trained for it.

Licensing model¶

You do not need a separate listing per rank. Publish the subject once; a buyer then configures a license along two independent axes:

• Role — the scale of the part the LoRA plays in a shot.
• Usage — how the buyer intends to use the output.

The buyer selects one option from each axis, and the combination sets the price (you set the underlying prices — see Pricing, payouts, and licensing). The Role tiers are read from the ranks you attach to the listing, so a single listing can serve both a background extra and a hero close-up. At purchase the buyer also picks which base model version they want (see One subject, multiple base models).

Role — what scale of part this is for:

Usage — what the output will be used for:

Because Role and Usage are independent, the same artifact can be licensed for a hobby project or a broadcast production without separate listings.

Build new variants on the platform engine¶

You can train new variants on the platform engine without local or cloud GPU. You keep the resulting weights and the same 85% revenue split applies.

The flow has four steps:

1. Upload your dataset. Images, captions, tags, and reference frames are all accepted.
2. Configure the run. Set the base model, rank, learning rate, step count, and number of variants. Every combination you specify is run.
3. Receive the variants. A single run can return multiple LoRAs — for example ranks 16/32/64 at different strengths and recipes, up to eight LoRAs from one dataset.
4. List the keepers. Preview each variant, then publish the subject and attach the ranks you keep as license tiers.

The "up to eight per run" output maps directly to the multi-license listing model above: one dataset and one run can produce the full Crowd-through-Marquee ladder for a single listing.

Dataset size and training time¶

Steps-to-convergence is driven by rank, not dataset size, so a larger dataset does not increase training time. With more images, each image is seen fewer times, which improves generalization. A varied 150-image set generally outperforms a thin 50-image set at the same rank. Prioritize capture variety before increasing rank.

How rank maps to a role¶

Lookhouse organizes LoRAs by role — how large a part the LoRA plays in a shot. Buyers browse and license by role, and role is read from the rank of the artifact. Because the three LoRA types do different amounts of work, the same role corresponds to a different rank for each type.

When choosing a target rank, keep three points in mind:

DoRA delivers comparable quality at a lower rank. DoRA at rank r typically matches plain LoRA at rank 2r in quality, with a smaller weight footprint, so it composes more cleanly when a buyer stacks it with other LoRAs. Use it for premium tiers.

Rank 128 is rarely the right choice. For a character, use DoRA at rank 32. For a large location, split it into Hero-tier zone LoRAs and bundle them. For a large transition, ship two rank-64 packs instead of one rank-128. Splitting generally produces better results than increasing rank.

Test composition before listing. Buyers commonly load a character, a location, and a transition together. Test your LoRA against representative LoRAs of the other types before listing, and document the working weight ranges in the listing. If a morph needs 0.7 weights rather than 1.0, note that — buyers need accurate ranges to avoid blurry first generations.

Pricing, payouts, and licensing¶

You choose free or paid for each LoRA and can mix tiers across your library. The paid split is flat at 85% to you, 15% to the platform. The platform does not adjust your set price.

Details:

• You set the price for every paid LoRA. There is no listing fee in either tier.
• The 85% split is flat and applies to every sale regardless of volume. There are no tiers to reach.
• Payouts run weekly via Stripe with no minimum balance; any amount can be withdrawn.
• Free listings are a valid strategy. An open-source Background LoRA can serve as an entry point to paid Lead and Marquee versions of the same subject.
• Price on dataset quality and license terms, not compute cost. Transition LoRAs cost the most to produce and typically command several times the price of a lead character LoRA.

Frequently asked questions¶

Do I keep ownership of my LoRA? Yes. Your weights remain your IP. The platform hosts and sells them on your behalf and does not claim them or train on them.

Can I list here and stay on Hugging Face or GitHub? Yes. Free and Standard listings remain live wherever else you publish. (Exclusive options are planned; early uploaders get first access.)

What base models are supported? SDXL, Flux, SD 1.5, Pony, and custom checkpoints.

Do you train on my LoRA or my dataset? No. The build engine trains only the variants you configure from your dataset and does nothing else with them.

How do payouts work? Through Stripe, weekly, with no minimum. Any balance can be withdrawn.

What if I want to leave? You can remove any LoRA from the marketplace at any time. Buyers keep what they have already purchased; new sales stop. There is no lock-in.

Get started¶

If a LoRA is already trained, it can typically be listed in about 15 minutes.

• Upload your first LoRA →
• Open the build engine →

Related: How licensing works (buyer view) — what a license grants and how role maps to part size.