Home › Datasets › Character & World Consistency
AI Filmmaking PackHow to Fine-Tune an LLM for Character & World Consistency
Building and maintaining consistent characters and environments across AI-generated images and video — reference sheets, identity locking, and drift troubleshooting.
- 730 instruction-response pairs
- Format JSONL
- Pack AI Filmmaking
- Dataset $39 one-time
$39 for the dataset. Training is priced separately by model, starting at $47.
What can this kind of fine-tuning teach an LLM?
A general-purpose model may already know some of this subject. Fine-tuning serves a different purpose: it repeatedly demonstrates the kinds of requests, response patterns, terminology, decisions and workflows you want the model to handle more consistently.
For Character & World Consistency, the dataset focuses on building and maintaining consistent characters and environments across ai-generated images and video — reference sheets, identity locking, and drift troubleshooting. Fine-tuning shapes recurring model behavior; it does not turn the model into a live database or guarantee correctness on every new case.
What training data should cover
Useful supervised fine-tuning data should resemble the work you expect the model to perform. This Rebiha dataset's domain definition is organized around 16 coverage areas:
- building a character turnaround sheet with front, three-quarter, profile, and back views composited into one reference; diagnosing why a character's face changes when the camera angle changes; using first-frame and end-frame generation to control a clip's start and landing state; deciding between frame-chaining and first/end-frame generation for a sequence; writing concrete edit instructions that preserve identity versus subjective ones that cause drift
- locking wardrobe and props separately from facial identity across generations; building a multi-angle reference set instead of relying on a single portrait; troubleshooting output drift that appears past a model's natural consistency window; chaining short clips with a last-frame hand-off for a sequence longer than one generation; anchoring multiple characters independently within a single multi-character scene
- maintaining a consistent environment or setting across separate, unrelated generations; building a reusable style lock for consistent art direction across an entire project; diagnosing the most common root causes of character consistency failure; using structural conditioning like pose or depth control to preserve identity across different poses; compositing multiple character references onto one canvas for parallel batch generation
- documenting a character's specific physical attributes in a reusable reference sheet for a project
How Rebiha prepared the training material
This page contains 730 instruction-response pairs from the existing Rebiha dataset. Its domain definition was designed around the coverage areas above so the training material spans multiple realistic tasks rather than one narrow prompt pattern.
Rebiha's broader dataset workflow uses structured quality review to help identify problems such as contradictions, unsupported claims, unsupported assumptions, calculation errors and other weak training examples. You can inspect real examples on this page before deciding whether the dataset fits your model and use case.
What is in this dataset
730 instruction-response pairs covering building and maintaining consistent characters and environments across ai-generated images and video — reference sheets, identity locking, and drift troubleshooting. Each pair is a realistic instruction with a full written response, in the JSONL format used for supervised fine-tuning. No synthetic filler, no duplicated rows.
Example pairs
Real rows from the dataset, shown in full.
How do I build a character turnaround sheet for my AI image generator so I can keep the same face across multiple shots?
Generate a single image containing four views of the same character: front, three-quarter left, profile left, and back. Use a prompt like 'character turnaround sheet, same person, identical face and hair, front view, three-quarter view, profile view, back view, neutral expression, plain gray background, full head and shoulders, consistent lighting, 4-panel grid'. If your tool can't do a grid in one pass, generate four separate images but force identity by reusing a fixed seed, adding 'same character as previous image' with a detailed face descriptor, and keeping the same lighting direction. Then composite the four into one reference image using an image editor, labeling each quadrant. Use that composite as your character reference for every subsequent generation—most tools accept it as an input image and will lock onto the facial structure. For best results, include a close-up of the eyes and a note on distinctive marks (e.g., 'small scar on left eyebrow') in the prompt alongside the reference.
My character's face subtly changes every time I change the camera angle from front to side view, even when I use the same seed. What's causing this and how do I fix it?
This is classic identity drift caused by the model interpreting 'side view' as a new character unless you anchor identity separately from pose. The fix has three layers: (1) Start with a character reference image—use your turnaround sheet or a single front-facing portrait—and pass it as the image prompt/reference for every angle. Don't rely on seed alone; seeds don't guarantee identity across different compositions. (2) Add explicit identity-locking language to your prompt: 'same person as the reference image, identical facial structure, same eye spacing, same nose shape, same mole placement, only change is camera angle to profile'. Avoid words like 'different angle' without 'same person' reinforcement. (3) If drift persists, generate an intermediate three-quarter view first, then derive the profile from that three-quarter result using an img2img workflow with low denoising strength (0.3–0.5). That forces the model to morph the pose rather than reimagine the face. Also check whether your tool has a 'face lock' or 'identity preservation' slider—if yes, set it to maximum. Finally, diagnose by generating the same side view ten times and comparing eye-to-eye distance and jawline; if they vary, your reference image lacks distinctive high-contrast features, so add a unique accessory (e.g., a specific earring) that appears in all angles.
I'm generating a 5-second video clip of my character walking from left to right. How do I use first-frame and end-frame generation to control exactly what the start and end look like?
First-frame/end-frame generation works by giving the video model two static images: one for the clip's opening frame and one for its final frame. The model interpolates the motion between them. Step 1: Generate your start frame—a full-body shot of your character at the left edge of frame, facing right, mid-stride. Use your character reference to ensure identity. Step 2: Generate your end frame—the same character, same outfit, at the right edge, but with a different leg position (e.g., opposite stride phase) and slightly shifted background if needed. Crucially, both frames must share: identical face, hair, clothing details, and lighting color temperature. If they differ, the model will morph or 'pop'. Step 3: In your video tool, load start image and end image, set duration to 5 seconds, and set motion strength moderate so the model doesn't overshoot. To avoid a jarring jump, ensure the end frame's pose is a natural continuation of the walk cycle—if your start has left leg forward, your end should have right leg forward or a neutral stance. If your tool doesn't accept an explicit end frame, you can generate the video from the start frame and use a prompt like 'walks to the right, ends with right foot forward', but explicit end frames give tighter control. After generation, check the last 5 frames for identity consistency; if the face drifts, regenerate the end frame with a lower denoising strength from the start frame.
Who this is for
Builders who want their fine-tuned model to reliably perform this specific capability, as part of a broader assistant or agent.
Which models this works with
The dataset is plain JSONL, so it works with any instruction-tuned open model. On Rebiha you can train it directly on:
- Qwen — Qwen3.8, Qwen3.6, Qwen3.5, Qwen3, Qwen2.5, Qwen2.5-Coder
- Granite — Granite 4.1 (3B, 8B, 30B)
- OLMo — OLMo 3, OLMo 3.1
- GLM — GLM-4
- Gemma, Phi, DeepSeek, Mistral, and Llama families
Training uses LoRA, or QLoRA on a 4-bit base for larger models. Your base model's weights are never modified — training produces a separate adapter.
How it works
- Pick this dataset and a base model.
- Rebiha runs the job on a managed GPU. You set up nothing.
- Download your model: a ready-to-run GGUF, plus developer assets (adapter, tokenizer, configs).
Fine-tuning or RAG?
Use RAG when the model needs facts that change — your prices, your documents, today's inventory. Use fine-tuning when you want consistent behavior and structure without prompting for it every time. This dataset teaches the shape of this capability, not facts to look up.
How managed fine-tuning works
- Inspect the dataset examples and choose a compatible base model.
- Rebiha runs the training job on managed GPU infrastructure using LoRA or QLoRA as appropriate.
- Download portable model assets for your own workflow, including the available ready-to-run and developer outputs for the selected training path.
Fine-tuning or RAG?
Use fine-tuning when the objective is recurring behavior: response structure, domain patterns, terminology, workflow, or how the model approaches a class of requests. Use RAG when answers depend on changing or proprietary facts that should be retrieved at request time. The two approaches can also work together.
Know what goes in. Own what comes out.
You can inspect real examples before training, choose the model you want to adapt, and keep portable model assets when training is complete.
Related datasets
- General Instruction Following — 978 pairs
- Multi-Turn Conversation — 466 pairs
- Rewriting & Editing — 854 pairs
- Summarization — 781 pairs