Reference Workflow

Early-access creators demonstrated MiniMax H3 Image generating Western game scenes, fantasy environments, and character sheets. A separate test assigns distinct references to character, setting, and lighting.

Early-access testers show MiniMax H3 Image reusing characters, preserving vehicle designs across views and changing artistic styles. They also report support for 16 reference images and 4K output.

Creators demonstrate Claude-built Blender scenes as references for Seedance 2.5 video generation. GlennHasABeard details Firefly's full-length reference input and warns that detailed previs can leak into the output.

Rahul’s open-source Breakboard collection uses manufacturer drawings, CAD and three illustration skills to build interactive Three.js teardowns. The standalone HTML runs offline in browsers that support WebGL 2.

Hasantoxr shared tests in which Tencent Hy4 turns sketches into games or HTML pages and app screenshots into code. The thread includes WorkBuddy access steps, API pricing and a November 15 cutoff for the limited free quota.

MengTo shared a workflow that turns a Figma Weave canvas into a seasonal landing page. Seasonal images, depth maps and Kling motion references feed an exported folder that Claude Code converts into a single-file three.js page.

Bilawal Sidhu demonstrates using GPT-6 Astra to turn a RealityScan room mesh into an editable Blender scene with isolated objects and procedural materials. He also demonstrates physics in the rebuilt scene and distinguishes it from point clouds or an unstructured triangle mesh.

MengTo built eight landing pages for about $12 in tokens using screenshots and revision prompts. The workflow used Nano Banana photos and Claude Code fixes; five of 24 first deliveries needed no repairs.

MengTo’s Dola workflow turns a product description and screenshot into campaign concepts, posters and animated clips. It uses Seedream 5.0 Pro for posters with real UI and Seedance 2.0 for motion.

Magnific says its short film SMOKE used character sheets and location references to maintain continuity across 42 shots. Production took ten working days, all video was generated with Seedance 2.5, and published prompts detail five locations.

Gizakdag and Egeberkina demonstrate staging scenes in CineMap before generating video in TapNow. Their examples use reference images or Seedance 2.5 to turn directed previews into finished shots.

Gokayfem used Claude Opus 5.5 to reconstruct a garden from a Gaussian splat in Blender. The open-source project combines procedural objects, Geometry Nodes, Shader Nodes and captured textures.

Venturetwins reports a consistent Kling 4 tour generated from eight listing photos in one pass. GPT-6 Astra selected the references and wrote a timed script, and the test used Kling’s built-in speech.

MengTo built landing pages with MiniMax M3.1 Flash, using images to set visual direction before coding. Shared references, reusable skills and screenshot reviews guided revisions.

Amir Mushich built an Off-Script coffee video using vector brand assets and 13 directed keyframes. His workflow uses Seedance through Lumina for generation, followed by editing, timing, sound and color work.

Curious Refuge’s tutorial turns greybox footage of an actress drinking coffee into a bar scene. Seedream sets the look, Depth Anywhere extracts depth guidance, and Seedance 2.5 generates the final video.

Amir Mushich used Revolut’s website and Claude to create an unofficial brand guide with HTML and design tokens. He says the results still need human checking.

Adobe ambassador CharaspowerAI used Firefly Boards to compare logo concepts and refine staff character sheets for a haunted hotel campaign. The workflow established consistent references before video production.

Figma Specs brings collaborative markdown onto the canvas in open beta. MCP and GitHub connections move context into and out of Specs.

MayorKingAI publishes the prompts behind Rama vs Ravana. The workflow uses GPT Image 2.5 character sheets, a Midjourney 8.2 battlefield, and a Seedance 2.5 battle prompt with timed cuts and reference-preservation instructions.

Erik Luo demonstrates Atlas turning Midjourney 8.1 images into textured, rotatable 3D characters and a stylized room. He reports better output from AI-generated inputs.

Kitze launched Flavors with copy-paste landing-page references, design Markdown, shadcn components and templates. Most resources are free; a launch-week $99 lifetime Pro plan adds API/MCP access and other assets.

Venturetwins reports making nine successive FLUX 3 edits to a Chloe ad without visible degradation. Magnific offers the model with multiple image references, element placement controls and targeted detail editing.

FLUX 3 Image adds bounding-box layouts, precise multi-turn edits, 4K output and support for up to 10 references. A creator comparison found it most accurate at following fine-detail editing instructions.

Magnific used Ideogram 4.5 to create cast and apartment reference sheets for a generated sitcom scene. After a video frame showed the wrong T-shirt, the team edited the frame and regenerated the clip.

Kling opened 4.0 Flash to Ultra Yearly subscribers ahead of the planned October launch of flagship 4.0. Early creator tests show text-to-video clips and scenes using character references.

Amir Mushich reports using GPT-6 Astra and Unreal Engine 5.8 to turn 20 real-estate listing photos and map panoramas into an interactive property walkthrough. He says Blender tests did not work well for the project.

Higgsfield demos show Claude Opus 5.5 converting paintings, drawings, character sheets, floor plans, and photos into editable Blender-style scenes. One demo also builds a browser app for placing photo-derived furniture in an interior.

A creator used Astra to turn an image mockup into printable papercraft instructions for a foldable pet toy. The workflow asks Astra to check and revise the design so the finished template can physically fold.

Creators are using explicit stillness constraints and named positions in Seedance 2.5 prompts to control cat behavior and head counts. In one 15-second test, a five-cat scene repeatedly returned four until each cat had a distinct position.

The workflow locks a realistic dashcam-style opening frame in GPT Image 2.5, then animates it in Seedance 2.5. Its prompts specify camera framing, branded vehicles, and escalating visual interruptions.

The workflow gives GPT-6 Astra authenticated MCP access to Hyper3D Rodin for planning and generating 3D assets. A related setup sends an image through Astra and Codex to create an editable 3D space rather than a static model.

Creators are using stills, simple Blender previs, and reference clips to steer composition, character appearance, materials, and camera framing in video models. The examples span Seedance 2.5, Hailuo, and MiniMax H3.

A creator says GPT-6 Astra recreated a studio as a Blender scene from five photos and three panoramas in 11 minutes. The model and web viewer were published for inspection, while a separate demo showed a similar office.

Higgsfield demoed GPT-6 Astra turning visual references and plans into editable Blender models, including nail art, an AirPods case, a Mies house, and Space Shuttle Discovery. The shuttle demo contains 192 assemblies.

Creators report using GPT-6 Astra, reference materials, and Blender to build editable 3D models. Reported results include an 11-minute studio model, a 3D-printable replacement part, and a cleaned RealityScan mesh.

Higgsfield says Astra converted a floor plan into scene code, a Blender location model, and a Cycles render. Separately, a creator says it made a to-scale cabin walkthrough from a 38-page PDF in about 10 minutes.

Magnific says MiniMax H3 Max can use up to 12 image, video, or audio references for clips up to 15 seconds. Its examples separate character, wardrobe, expression, and location references to maintain continuity across shots.

In a paid Adobe partnership, a creator documented a trainer-sandal campaign made in Firefly Boards. The workflow moved from product concepts and material boards to lifestyle images and motion clips using the sandal as reference.

Midjourney updated its V8.2 Edit test model with instruction edits, up to four image references, 2x --hd output, inpainting, and outpainting. Midjourney says the update improves image quality and is seeking feedback.

Practitioners are using MiniMax H3 Max with start-end frames, image-to-video, and text-to-video inputs. Community guides also cover H3 reference-to-video nodes and motion-control timelines for character work.

Gemini Omni Flash 1.1 adds video references, 10-second extensions, and first- and end-frame controls. It offers 360p drafts, 4K upsampling, and access through Pika API Club.

Creators report using up to six image references and short clips in MiniMax H3 for reference-led scenes. One workflow used the setup to make a Lego-style sitcom sequence for a presentation.

Magnific says creators can set start and end frames before generation to produce scene changes in one run. The workflow numbers and timestamps shots in the prompt while holding reference images.

Creators are using image references to place prepared characters in generated scenes and guide image-to-video shots. Higgsfield's Seedance 2.5 workflow also swaps people, backgrounds, or effects from supplied photos.

Magnific demonstrates a Wan 3.0 clip guided by one character sheet and a prompt specifying retained identity traits, scene direction, camera, and audio. Early creator tests say referenced styles can still shift toward a 3D look.

Creators describe using storyboards or generated stills as first, last, and reference frames for MiniMax H3 shots. One demonstrated the approach with a timed racing sequence, then assembled the results in conventional editing tools.

Creators use Seedance 2.5 reference assets to preserve character identity and spatial continuity across long sequences. One workflow turns two initial character frames into seven 30-second generations for a 3.5-minute film.

A creator used MiniMax Design to analyze two meme references and plan an FPV-style transition between them. The agent produced an H3 prompt specifying opening and closing frames for one continuous move.

Creators are using character sheets, prior clip endings, and anchor frames to maintain identity, lighting, and shot continuity across generated video. Reported workflows pair reference assets with Seedance and game-camera setups.