CHAAD Forge

Approved benchmark labels

Loading approved benchmark labels…

Recommendations from your validation set

Select a fully approved benchmark to see recommendations based on its validation frames and saved boxes.

Pipeline settings

Choose what to generate and how it should match your real examples

Dataset mode

Generation profile

Client environment

Use an accepted clean plate to preserve a client camera view. Without one, generation creates a new scene from the prompt.

Manage backgrounds

DreamBooth adapter

Reuse a private subject or product adapter. Client locations belong in Backgrounds, not DreamBooth.

0 references ready · minimum 3 · recommended 20–50 diverse images

Train DreamBooth separately

Loading saved models…

Upload client references

Training scene composition

Match the camera view, worker scale and activities in your approved validation frames. Use the recommendations above as a starting point, then adjust for scenes the detector needs to learn.

Scene planning

Automatic planning uses the source camera and metric depth to choose worker count and scale. If forklift is a dataset class, applicable images place one measured forklift with one worker; the seated-operator pose targets its actual cab.

Routine worker limitRare five-worker scenes

Camera angle

Camera distanceBody posePeopleChosen automatically from scene capacity.

Human appearance

Broad variation changes adult appearance and clothing across training scenes. Use prompt-only when exact attributes are specified. Spotters appear only when requested; a nearby person is not automatically a spotter.

Training workflow: generated scene → synthetic examples of your selected classes → annotation proposals → quality review.

Generated dataset size1,000 images

Separate from the DreamBooth reference count. Cost and runtime are reviewed before paid GPU work starts.

Annotation proposals use the pipeline threshold internally. Saved, human-approved boxes define the benchmark labels.

Seed images

captured uploads

Pipeline mode

both

annotation / synthetic

Split

TRAIN

target cohort

Image augmentation

Apply rotation, flips, color science, and lighting jitter to create harder negatives.

Text prompts (generation)

Dataset classes

Labels and review

Annotation proposals use the selected class list. When a benchmark is selected, its approved label definitions are shared by generation and annotation.

Snapshot

Pipeline: direct synthetic generation
Target dataset size: 1,000 images
Human review pauses at 10, 40, and 100 generated images (up to your target). Review labels, then explicitly continue each checkpoint.
Augmentation: enabled
Client appearance adapter: off
Scene-native pose: mixed · medium · automatic, usually up to 4 people (3% five-person scenes)