Gemini Multi-Turn, FLUX LoRA, and Cross-Platform Consistency

Character DNA Prompt Technique

The "Character DNA" technique front-loads a comprehensive character description at the start of a Gemini multi-turn session, then references it as a persistent constraint:

Example: Captain Eva Rostova

Turn 1 — Establish the DNA:

I'm creating a consistent character. Here is her Character DNA — maintain
these exact features in every image you generate in this conversation:

CHARACTER DNA:
- Name: Captain Eva Rostova
- Age: 38
- Ethnicity: Eastern European (Russian)
- Face: High cheekbones, angular jaw, strong brow, slight aquiline nose
- Eyes: Steel grey, deep-set, intense
- Hair: Dark auburn, shoulder-length, usually pulled back in a low bun
  with loose strands framing face
- Build: Athletic, 5'10", broad shoulders, lean
- Signature wardrobe: Charcoal military-inspired coat with standing
  collar, black fitted base layers, leather utility belt
- Distinguishing marks: Small scar through left eyebrow, silver ring on
  right index finger
- Expression default: Controlled intensity, rarely smiles, when she does
  it's subtle and asymmetrical

Generate a portrait of Captain Eva Rostova, medium shot, standing at
the bridge of a spacecraft, looking out a viewport. Soft blue ambient
light from control panels. Shot on Hasselblad, 85mm, f/2.

Turn 2 — New scene, same character:

Now show Eva in a completely different setting: sitting in a dimly lit
bar on a space station, nursing a drink. Leather jacket instead of
her military coat. Low key lighting, neon accents. Same face, same
scar, same ring. Shot on Leica M11, 35mm, CineStill 800T.

Turn 3 — Action scene:

Eva in combat. Running through a corridor, coat flowing, pulling a
sidearm from her belt. Motion blur on the background, sharp focus on
her face. Intense expression. 24mm wide angle, dramatic rim lighting.

Critical Reminder Technique

In longer sessions, Gemini can drift. Periodically reinforce the DNA:

REMINDER: Maintain Eva's exact features — steel grey eyes, auburn hair
in low bun, scar through left eyebrow, silver ring on right index
finger. Do not alter her facial structure.

Now generate: [new scene description]

FLUX LoRA Training

FLUX LoRAs (Low-Rank Adaptations) let you fine-tune a model on a specific character, style, or concept for near-perfect consistency.

Requirements

• GPU: Minimum 24GB VRAM (RTX 4090 or cloud equivalent)
• Training images: 15-30 high-quality images of your subject
• Software: FluxGym, Kohya_ss, or cloud training services
• Time: 30-90 minutes depending on steps and hardware

FluxGym Training Process

1. Collect training images — 15-30 images of your subject. Variety is key: different angles, expressions, lighting, backgrounds. All should be high quality and clearly show the subject.
2. Prepare captions — Each image needs a text caption describing what's in it. Auto-captioning tools can help, but manual review improves results.
3. Configure training — Set learning rate (typically 1e-4 to 5e-4), training steps (500-1500), batch size, and resolution (typically 512x512 or 1024x1024).
4. Train — Run the training process. Monitor loss curves — you want steady decrease without overfitting.
5. Test the LoRA — Generate test images at various weights (0.5 to 1.0) to find the sweet spot.
6. Deploy — Use the LoRA file in your generation pipeline. Combine with text prompts for scene control.

Key Training Tips

• More diverse training data = better generalization. Don't use 20 photos from the same angle.
• Caption quality matters as much as image quality. Bad captions = bad LoRA.
• Train at the resolution you'll generate at. Mismatch causes artifacts.
• Don't overtrain. If your LoRA only produces images that look exactly like training data regardless of prompt, reduce steps.
• LoRA weight at generation time is crucial: too low and the character won't appear, too high and the image loses flexibility. Start at 0.7.

Flora Character Reference Node Workflow

Flora AI provides a node-based system for character consistency. Here's how to set up a character pipeline:

5-Step Process

1. Create a Character Reference node — Upload 3-5 reference images of your character to the node. These establish the identity.
2. Connect to a Generation node — Wire the Character Reference output to a Generation node's reference input.
3. Add a Prompt node — Write your scene description. Connect it to the Generation node.
4. Set consistency strength — Adjust the character reference weight (similar to Midjourney's --cw). Higher = more consistent face, lower = more creative freedom.
5. Generate and iterate — Run the pipeline. If consistency is too loose, add more reference images or increase weight. If the character looks stiff or "pasted in," decrease weight.

Flora's advantage is that the pipeline is visual, repeatable, and shareable. Once you build a character consistency pipeline, anyone on your team can use it.

Exercise

Cross-Platform Character Challenge

1. Design a character with detailed physical attributes (use the Character DNA format).
2. Generate 3 images of that character in Gemini using multi-turn editing with the DNA technique.
3. Generate 3 images of the same character in Midjourney using --cref or --oref.
4. If you have access, generate 3 images using a FLUX LoRA or Flora pipeline.
5. Place all 9+ images in a grid. Evaluate:
   • Which platform maintained the most consistent identity?
   • Which offered the most creative flexibility?
   • Where did each platform struggle?
6. Document your findings — this comparison will inform your production workflow decisions.