Higgsfield claims GPT-6 Astra created 120-frame butterfly animation with 600 editable Photoshop layers

• GPT-6 Astra reportedly worked inside a Photoshop workflow for five to six hours, producing 120 butterfly frames and 600 editable layers, according to Higgsfield's main post and Higgsfield's final frame count.
• Higgsfield described a four-part pass: generate a flight sequence, supply handmade-animation references, make charcoal-and-pencil brushes, then draw the frames layer by layer in Higgsfield's first step.
• The material references were meant to teach paper and drawing behavior, while Higgsfield's brush step says Astra made custom brushes to carry charcoal and pencil texture through the work.
• The same account has also shown rough drawings being refined into smoother animation by GPT-Image 2.5 in Higgsfield's rough-sketch demo.

A September report describes a separate Higgsfield test in which GPT-6 Astra and Fable 5.1 drew Van Gogh in Photoshop by selecting their own tools. The butterfly experiment is more ambitious in duration: Higgsfield's final update puts the run at 120 frames rather than a single canvas.

Higgsfield says Astra added layers step by step over five to six hours while preserving paper texture and variation in each stroke. Its reported deliverable was a Photoshop file with 600 editable layers for the 120-frame sequence in Higgsfield's final update.

The posts do not provide the source file or a layer inventory, so the reported counts are Higgsfield's characterization of the run.

Higgsfield's thread lays out the production sequence:

1. Draw a butterfly in flight as a sequence of images with GPT-Image 2.5 in Higgsfield's first step.
2. Feed Astra handmade animations to model the behavior of paper and drawing materials in Higgsfield's reference step.
3. Create custom Photoshop brushes that imitate charcoal and pencil as Higgsfield reported.
4. Draw the 120 frames incrementally, retaining texture and stroke variation in Higgsfield's final update.

The handmade-reference step is the notable constraint. It describes a workflow where material behavior is supplied before the long drawing pass, rather than treated as a final surface effect.

Rough sketches can serve as motion scaffolds: Higgsfield says GPT-Image 2.5 enhanced frames, refined them, and smoothed movement from a hand-drawn animation in its demo.

Underwoodxie96 wrote that separate poses had made their sprite-sheet animations choppy. Their fix was the instruction, “All 16 frames must form one continuous sequence of the same action,” within a prompt that specifies a 4-by-4 grid, frame order, consistent scale and alignment in the full prompt.

Higgsfield also isolated continuity as its own test: “A new generation shouldn’t mean a new object or position.” Its stop-motion clip tests consistent illumination across GPT-Image 2.5 frames in Higgsfield's post.