Which Board Features Work With Each Model?

When you attach a moodboard to a generation in RandomSeed, something happens under the hood — but what exactly depends on which model you have selected. FLUX Dev applies your board's LoRA adapters. Ideogram v3 uses your board images as style references. FLUX Pro Ultra uses a selected board image as a seed. And every model, regardless of tier, receives prompt fragments distilled from your board's visual analysis. This guide maps out the full matrix so you can choose the right model for the kind of influence you want.

The Four Ways a Board Influences a Generation

RandomSeed boards can influence image generation through four distinct mechanisms. Each operates at a different level of the pipeline:

• Prompt fragments — A compact style description is derived from your board's centroid and appended to your prompt. Works with every model. Always active when a board is attached.
• LoRA conditioning — LoRA adapters trained on your board images are applied to the generation pipeline. Deep style transfer at the model weights level. Available for select FLUX models.
• Style Reference — Board images are passed as image conditioning inputs, guiding the model toward a visual style without modifying the generation prompt. Available for Ideogram v3.
• Seed Picker — You select a specific board image to use as an image-to-image starting point. The model generates a new image that shares the structural qualities of the seed. Available for models with i2i support but without LoRA.

Prompt Fragments — All Models

Prompt fragments are the baseline layer of board influence. When you attach a board, RandomSeed analyzes its images and produces a compact style description — typically covering color palette, mood, texture qualities, and compositional tendencies. This description is appended to your prompt automatically.

Fragments work with every image model because they operate entirely at the prompt level. They require no special model capability — just text conditioning, which all models support. If you're using a model that doesn't support LoRA, Style Ref, or Seed Picker, prompt fragments are still doing meaningful work.

LoRA — FLUX Dev, FLUX Pro 1.1, FLUX 2 Pro

LoRA (Low-Rank Adaptation) is the deepest form of board influence available. Rather than describing your style in text, LoRA adapts the model weights themselves to reflect the visual characteristics of your board. The result is a generation that has genuinely absorbed your aesthetic — not just been told about it.

In RandomSeed, LoRA conditioning is available for FLUX Dev, FLUX Pro 1.1, and FLUX 2 Pro. When you select one of these models and attach a board, the LoRA adapters associated with your board images are applied automatically. You don't need to configure anything — it runs in the background.

For a detailed explanation of how LoRA compares to prompt-only style transfer, see LoRA vs. Prompts.

Style Reference — Ideogram v3 Only

Style Reference uses image conditioning to guide the model toward a visual style. Your board images are passed directly as reference inputs, and the model uses them to steer the aesthetic of the output — without the LoRA adapter machinery.

Currently, Ideogram v3 is the only model with Style Reference support. FLUX models have an equivalent API capability (reference_image_url and ip_adapters), but the fal.ai pipeline for those parameters is currently broken and has been disabled. FLUX boards still receive prompt fragments and, where applicable, LoRA conditioning.

Seed Picker — FLUX Pro Ultra, Recraft V3, Grok Imagine

The Seed Picker gives you direct control over image-to-image conditioning. You select a specific image from your board, and RandomSeed uses it as the starting point for the generation — at 0.75 strength by default, meaning the output shares the structure and color feel of the seed while still being a new image driven by your prompt.

Seed Picker is available for FLUX Pro Ultra, Recraft V3, and Grok Imagine — models that support image-to-image but not LoRA. For LoRA-capable models, the Seed Picker is intentionally hidden: LoRA provides a stronger and more consistent form of style conditioning, so running both simultaneously would produce conflicting signals.

Feature Availability by Model

Choosing the Right Model

The best model depends on how strongly you want the board to influence the output:

• Maximum style fidelity: Use FLUX Dev, FLUX Pro 1.1, or FLUX 2 Pro. LoRA conditioning produces the most consistent and deep style transfer. FLUX Dev is the best cost-quality balance; FLUX Pro 1.1 and 2 Pro offer stronger prompt adherence for complex descriptions.
• Distinctive aesthetic with style reference: Use Ideogram v3. Its style reference conditioning produces outputs with strong visual character that closely mirrors your board images. Ideogram also tends to render stylized illustrations and graphic designs well.
• Compositional starting point: Use FLUX Pro Ultra, Recraft V3, or Grok Imagine with Seed Picker. Select a board image whose composition, layout, or color palette you want to carry into the new generation. FLUX Pro Ultra gives the highest output resolution; Recraft V3 is strong for graphic and design-oriented outputs.
• Prompt-driven with board context: Any model. Prompt fragments alone are a meaningful influence — if you want a model like HiDream Fast or GPT Image 1 for their specific characteristics, the board still informs the generation through style description in the prompt.

Attach Your Board and Generate

Build your moodboard in the moodboard studio, select the model that matches the conditioning approach you want, and generate in the canvas. The feature badges in the model picker show you exactly which board features activate for each model — so you always know what's running before you hit generate.