Track 3 — Full ImagenWorld (six tasks: TIG, SRIG, MRIG, TIE, SRIE, MRIE)
Goal
Cover all six ImagenWorld tasks: text-guided generation/editing plus single/multi-reference generation/editing.
Why this is a research project, not a notebook extension
Not feasible in 3–5 h of notebook training. The paper's evaluated models (SDXL, GPT-Image-1, FLUX, Qwen-Image, OmniGen2, BAGEL) are multi-GPU-week systems. To honestly appear alongside them requires:
- Reference-image conditioning (image encoder + cross-attention or concatenation).
- Editing conditioning (source image + instruction fusion).
- Multi-reference composition.
- Training data with all six task types, including editing pairs.
- ≥50 GPU-hours of training, likely far more.
Scope (sketch only, to be detailed when opened for work)
- Unified conditioning interface for
f(t_ins, [I_src], [I_ref1...], z).
- Multi-image encoder (e.g. DINOv2 or CLIP image encoder) for reference conditioning.
- Instruction-conditioned U-Net (cross-attention over fused text+image tokens).
- Editing loss preserving source structure (e.g. latent-mixing + reconstruction term).
- Dataset construction pipeline for six task×domain combinations from the paper.
- Human-annotated failure-mode tagging on a small slice (object-level + segment-level via Set-of-Mark).
Evaluation
- Full ImagenWorld rubric: all six tasks × four criteria × VLM slice + (small) human slice.
- Krippendorff α and Spearman ρ vs. the paper's reported ranges as a sanity check.
Risk
The risk in starting this under the "3–5 h notebook" framing is shipping six half-finished notebooks whose models are far from converged and whose outputs are uninterpretable blobs. If we actually want this, the time budget moves to ≥50 GPU-hours and we are building a small research project, not extending the tutorial. This issue is parked until Track 1 (and likely Track 2) are merged and a separate GPU budget is allocated.
Dependency
Blocked by Track 1 and Track 2.
Track 3 — Full ImagenWorld (six tasks: TIG, SRIG, MRIG, TIE, SRIE, MRIE)
Goal
Cover all six ImagenWorld tasks: text-guided generation/editing plus single/multi-reference generation/editing.
Why this is a research project, not a notebook extension
Not feasible in 3–5 h of notebook training. The paper's evaluated models (SDXL, GPT-Image-1, FLUX, Qwen-Image, OmniGen2, BAGEL) are multi-GPU-week systems. To honestly appear alongside them requires:
Scope (sketch only, to be detailed when opened for work)
f(t_ins, [I_src], [I_ref1...], z).Evaluation
Risk
The risk in starting this under the "3–5 h notebook" framing is shipping six half-finished notebooks whose models are far from converged and whose outputs are uninterpretable blobs. If we actually want this, the time budget moves to ≥50 GPU-hours and we are building a small research project, not extending the tutorial. This issue is parked until Track 1 (and likely Track 2) are merged and a separate GPU budget is allocated.
Dependency
Blocked by Track 1 and Track 2.