TL;DR

We're excited to announce full support for training custom models using Flux Pro and Ultra. Alongside this, we’ve rolled out a revamped Training Sets system, seamless S3 integration for uploads, improved status tracking, and tighter UI flows. Oh, and marketing’s been cooking too. 👀

Custom Training for Flux Pro & Ultra

The defining highlight this cycle is the ability to train high-performance custom models using our Flux Pro and Flux Pro Ultra tiers. These are purpose-built pathways for creators looking to push fidelity and control even further. The backend now distinguishes between standard and professional model types, applying appropriate routing, error handling, and inference APIs to each.

Users can now select fine-tuning types like lora or full, specify presets, and provide tuning descriptions. New progress tracking returns elapsed and remaining training time, with explicit guidance on how many images (three minimum) are needed to get strong results. Custom Forge workflows now dynamically adjust based on model type, fully supporting professional image creation and prediction status tracking.

Training Sets: Build Once, Train Many

To streamline repeated training workflows, we’ve introduced the concept of Training Sets. These are named collections of user-uploaded images that can be re-used across multiple models. Whether you're working on a style or a subject, sets can now be categorized, described, edited, and re-linked as needed.

Notable features:

• Fully integrated Training Set UI with previews, presets, and linked models

• Seamless support for transferring image thumbnails into S3 during creation for better performance and reliability

• Up to 800MB or 150 images per set—great for deep, rich training contexts

• Automatic zipping and upload to S3 using pre-signed POST URLs in the browser

These foundation blocks future-proof the way creators can organize and scale their AI training pipelines.

Upscaling Logic & Guidance Tweaks

Under the hood, we reposed the image upscaling system to intelligently apply 2x or 4x scaling depending on model type and settings—including correct treatment for ultra-quality images. Image guidance strength has also been refined, amplifying how closely outputs reflect your intended training likeness.

Improved Custom Model Creation Flow

The UI for creating custom models got a makeover. It now reveals step-by-step feedback, with real-time pricing, preset selectors for model type and tuning mode, and a more polished handoff from training set to model. Users are clearly informed about what each field does, making the whole process smoother and less error-prone.

Better Data Handling with S3

We’ve moved from local file processing to a mature cloud-based storage pipeline. All training images and their thumbnails are now uploaded and hosted via Digital Ocean S3, which enables better storage performance and secure read operations through signed URLs. Combined with new validation rules for images and file types, this significantly boosts reliability during model training.

Marketing & Growth Moves

On the non-dev side, we made key strides in our July marketing strategy. Here’s what’s happening:

• Technical SEO: Sitemap cleanup and robots.txt configuration now reflect new domains and content correctly. Expect stronger organic visibility moving forward.

• Outreach: KOL and Pinterest marketing are underway, with creator activation initiatives in active testing and planning phases. Our top users are offering feedback, and aggregated sites have started indexing our best ai art generator tools.

Also, look out for the upcoming update to the browse page, which will soon sort by popularity—making it easier to surface trending creations in the AI art playground.

Smaller but Sweet Improvements

• Training progress now shows precise duration info and lets users navigate away without disruption

• Thumbnails for training sets are larger, clearer, and generated faster

• Model cards now show the base model used (e.g., normal, advanced)

• Ahrefs analytics is re-enabled and properly scoped to the new site layout

What’s Next

Looking forward, we’re focused on rolling out:

• A Custom Model Browser to search and discover community-generated models

• Bug fixes for account and image linking issues

• Improvements to mobile navigation and homepage UX

Together, these updates lay the groundwork for a vibrant, creator-first ecosystem around custom ai art with custom models.

In Closing

Training a custom model should not feel abstract or inaccessible. With new training sets, S3-backed uploads, and deeper controls in Flux Pro and Ultra, you’re now equipped to build, refine, and evolve your visual style like never before. And we’re far from done. Stay tuned for more—because this playground is just getting started.

Try out the new Forge at app.fiddl.art/forge

And you can always follow us on the public Trello Board.