Skip to content
Claude CodedocsData & ML1 file · +97 −0

Regenerate Figure 2 ring pipeline diagram with PaperBanana

Session from traceweave on Hugging Face by Samarth0710 · MIT · condensed by Coders Talk

8msession
0interventions
GOAL

Improve the existing Figure 2 (ring pipeline diagram) in a Euro-Par paper by using the PaperBanana figure-generation tool to produce a better version.

  1. Prompt · +0
    I dont like fig 2, make it better using paperbanana
  2. Agent did · +0m
    Explored the paper's LaTeX source to locate Figure 2 (ring pipeline diagram) and investigated the PaperBanana tool's API and diagram-generation capabilities.
  3. Agent did · +2m
    Wrote a generate_fig2.py script to call PaperBanana in diagram mode with a prompt derived from the paper's methodology text, then ran it as a background task.
  4. Agent did · +7m
    Waited for the background generation task to complete; PaperBanana produced 3 refinement iterations of the new figure.
  5. Agent did · +8m
    Read and displayed all three generated iterations of the new Figure 2 to compare them.
  6. Fail · +8m
    API error: prompt token count exceeded 128000 limit while trying to finish selecting/summarizing the best iteration.
  7. Outcome · +8m
    Session ended with the new figure iterations generated but the agent unable to complete the comparison/summary due to a repeated token-limit API error.
VERDICT

Next time I'd tell it upfront to keep the image context small since loading multiple full-resolution figure iterations blew past the model's token limit right at the finish line. The PaperBanana generation itself worked fine and produced three usable iterations, but I never got the agent's final pick because of the context overflow. I'd ask it to describe or downsample the images rather than load them all raw for comparison.

I tried this · 0

Discussion · 0

How Builds work