Client work
Kingsbet · 2025
An Offer You Can't Refuse
Client work
Kingsbet needed a flagship key visual that radiated power and inevitability, the unmistakable presence of a boss who already knows the outcome. The creative answer was a deliberate nod to the most iconic crime boss in cinema, produced end to end through a brand-trained generative pipeline and shipped as a live campaign.
Creative direction, LoRA training, ComfyUI pipeline design, compositing and final grade.
The brief, decoded
The campaign brief asked for a single hero visual to anchor a high-stakes promotional push, an offer framed as something you simply do not turn down. The creative call was to lean on a cinematic icon everyone reads in a heartbeat: the composed, dominant crime boss who never raises his voice because he never has to.
The constraint was time. There was no window for casting, location scouting, or a studio day. Every deliverable had to be production-final within hours.
The idea and the obstacle
The creative solution treated the figure as a portrait of certainty: low, directional key light, a confident three-quarter pose, and an environment that recedes into brand-coloured shadow.
The technical challenge was consistency. A generic model produces a different face, wardrobe, and mood on every seed. The brand DNA had to live inside the model itself.
The pipeline, node by node
I trained a dedicated brand LoRA on a curated set of Kingsbet references so the model internalised the wardrobe, palette, and tonal restraint of the brand. That LoRA became the backbone of a custom ComfyUI graph that I drove like a camera rig.
Structure came from ControlNet; identity and material consistency came from IPAdapter; the final image was assembled from decomposed render passes so lighting and grade could be tuned non-destructively.
LoRA Dataset
- Trained domain
- Kingsbet brand wardrobe, palette, and tonal language
- Dataset
- 74 curated reference images
- Training
- 2,400 training steps, batch 4
- Base model
- FLUX.2 Kontext
ComfyUI Workflow
- 01 · Checkpoint Loader
- Base model: FLUX.2 Kontext base weights
- 02 · LoRA Loader
- Brand model: kingsbet_v3, strength 0.82
- 03 · CLIP Text Encode
- Conditioning: Positive / negative prompt pair
- 04 · ControlNet Apply
- Structure: Depth + pose, weight 0.65
- 05 · IPAdapter
- Identity: Reference embed, weight 0.7
- 06 · KSampler
- Sampler: dpmpp_2m, 32 steps, CFG 4.5
- 07 · VAE Decode
- Output: Latent to 4K render pass
Conditioning Notes
Depth ControlNet to lock the subject silhouette and spatial separation from the background
OpenPose ControlNet to hold the commanding three-quarter stance across seeds
IPAdapter reference pass to keep wardrobe and facial structure brand-consistent
Output Specs
- Lighting
- Single low-key directional source, hard falloff
- Color grade
- Kingsbet deep-navy and gold, crushed blacks
- Brand check
- Palette, wardrobe, and tone validated against brand system
Layer Decomposition
- Base render
- LoRA-driven subject pass at 4K
- Background plate
- Brand-shadow environment, separated for grade
- Key light pass
- Directional rim and cheek light, dodged in
- Material detail
- Fabric and skin micro-texture overlay
- Color grade
- Kingsbet palette LUT + film grain
Final Output
- Lighting
- Single low-key directional source, hard falloff
- Color grade
- Kingsbet deep-navy and gold, crushed blacks
- Brand check
- Palette, wardrobe, and tone validated against brand system