A 30-Second Anime Short Inside One Month of Firefly Credits

Most beginner guides to AI anime teach the pipeline and skip the invoice. CreativeEdge, a Japanese studio that produces a full AI anime series, went the other way round. Their part-one tutorial picks its tools specifically to fit inside the credits that already come with an Illustrator-class Creative Cloud Pro subscription, about 4,000 a month, and that constraint turns out to be the most transferable part of the tutorial.

The numbers decide everything else, so start there.

The price list that shapes the workflow

As of the author's September 2026 test, inside Adobe Firefly:

• GPT Image 2.5 at 1K, high quality: 5 credits. That is 800 stills on a monthly grant.
• GPT Image 2.5 at 2K: 15 credits, or 266 stills.
• Kling 3.0 Omni at 720p, 15 seconds: 450 credits.
• Google Veo 3.1 at 720p, 8 seconds: 400 credits, and 8 seconds is the ceiling.
• Seedance 2.5 at 720p, 15 seconds: 1,425 credits.

Read that as a ratio rather than a list. One fifteen-second clip at the premium end costs roughly as much as 285 approved stills. So the workflow that follows is not a stylistic choice, it is the only shape that fits. Iterate on pixels that cost five credits, lock the plate, and spend video credits exactly once per shot.

If you want the same logic at larger scale, our cost-lane routing piece and the MiniMax H3 anime production notes cover what changes when the budget is a real production line.

Everything happens on the board, except the video

The author runs the whole asset phase inside a Firefly board: a canvas where generations land as objects you can reference, reposition and re-scale. The loop is the same every time. Generate, drag the result onto the canvas, then right-click and resize it to its original size so the scale on screen matches the real file.

Two details in that loop are worth copying. Resizing to original size keeps you from judging a composition at the wrong scale, which is how people end up with a character who looks right at 30% and is cropped at full size. And keeping assets as board objects means the next generation can point at them directly, which is how you get the character sheet and background into the same frame later.

1. The character, with the negatives written out

The base generation is deliberately verbose, and it is verbose about proportions rather than adjectives. The subject is a stylised boy, roughly 2.7 heads tall, large head, small torso, narrow shoulders, thin short limbs, hands hidden behind the back, plain white background, full body in frame.

Then the prompt lists what not to draw: no typical oversized anime eyes, no soft rounded chibi proportions, no realistic limb thickness, no thicker legs, no taller proportions, no gentle picture-book rendering. Model at GPT Image 2.5 Sunburst, 1K, high quality, portrait.

This is the cheapest identity insurance in the whole pipeline. Five credits buys a reference that every later step is measured against.

2. The character sheet is a layout spec, not a description

The second generation references the first image and asks for a character setting sheet in the same style, same design, same colours, on a plain white background. The instructions are structural:

• Three columns. Left: front view, full body. Centre: back view, full body. Right column split into three equal rows for a front bust, a right-facing profile, and a three-quarter left face.
• Both full-body figures at identical scale and height, filling the frame top to bottom.
• Minimum spacing between figures, as little waste as possible.
• The right-column faces drawn large, filling each cell, with hair or shoulders allowed to run off the edge.
• No text, no logos, no frame lines, no ground shadows.

Landscape, 2K, referenced to the character image. The author flags that picking 2K sometimes shifts the aspect ratio, so check the result and crop if needed.

Why this matters beyond this tutorial: a sheet like this is what later turns into a LoRA dataset, a video reference, or a handout for a collaborator. We covered the same asset-thinking for AI characters in the character consistency system guide.

3. The background gets a technique spec too

The background plate is generated with no people, animals or text in it, and the prompt describes the medium rather than the scene: poster colour on paper, flat brush marks with unevenness left in, smooth brush-blurred sky gradient, clouds built from thick white paint, no outline drawing, highlights painted on afterwards, and explicit exclusions for photo realism and digital gloss.

The composition instructions are the quiet craft lesson. Eye level at a child's height, horizon about 40% up from the bottom of the frame, and a wide flat area of ground from the centre to the foreground reserved for the character to stand on. Warm afternoon slant light, warm sunlit surfaces, cool blue-violet shadows, earth-tone palette with the sky as the brightest element.

Painting the empty space on purpose is what separates a background plate from a landscape picture. A beautiful illustration with no standing room is useless as a plate.

4. Assemble the scene cheaply

Now the payoff. One line of instruction, "the boy is walking happily," both the character sheet and the background plate referenced, 1K resolution for a rough. The character lands on the reserved ground, in the right light, in the right scale.

If that still comes out wrong, the fault is almost always upstream in the sheet or the plate, not in this step. That is the benefit of a cheap assembly pass: it tests whether the assets were specific enough, and it costs five credits to find out.

Video: same scene, different tool, and a reason

The author tried generating the animation on the board and stopped. In their test, board video generation ran without audio at all, with no audio switch to turn on, and charged 300 credits for a 10-second clip while the prompt panel displayed 100. So they downloaded the approved character sheet and background plate and ran video in Firefly's web generator instead.

Settings that matter:

• Model: Kling 3.0 Omni, 720p, widescreen, 10 seconds
• Audio: on. 250 credits with audio, 200 with it off, and the test produced voice
• Reference: add the two images through "image," not "frame"

The prompt carries the performance, the line, and the camera:

The boy walks along happily, looking around and muttering that the weather is lovely and that he is getting hungry. The camera is positioned in front of him and moves backwards at exactly his walking speed, a tracking shot. The distance between camera and boy stays constant from start to finish, so his size and position on screen never change. No zoom, fixed focal length, camera height at his chest and level. No music. No BGM.

Two of those clauses are doing more work than the rest. Naming the shot and then describing its invariant, the distance never changes, is how you get a stable travelling shot instead of a slow drift in and out. And the explicit audio-off instruction is the difference between a dialogue line and whatever soundtrack the model felt like adding. The general version of this camera language is in our cinematic prompt breakdown.

Two joined clips produced the author's roughly 35-second test piece. That is the realistic output of this method, not an episode.

What I would take from this

Spend your iteration budget where the pixels are cheap, and treat the video credit as the only decision you cannot take back. Almost every beginner AI anime project that stalls has done it in the opposite order: a few expensive video generations first, then a scramble to justify them.

Two more habits from the source that generalise. First, when a reference has to carry identity across many steps, specify its structure, not its vibe: the column and row layout is doing the heavy lifting, because a model can only hold the decisions you actually wrote down. Second, verify the tool's own meter. An account charged several times the number the panel displayed, plus audio that silently failed to generate, is a reminder that the UI is a claim and the ledger is the fact.

What to check before you trust these numbers

The credit prices here are from September 2026 and depend on the plan and the subscription tier, so re-check yours before planning a shoot. The tutorial is beginner-oriented and supplies its prompts outright; the deeper material, including the project files and the rest of the series, sits behind a paid membership.

The other limit is the one the budget sets. Four thousand credits a month produces short pieces with lots of still iteration and very few video renders. If your target is minutes rather than seconds, you are shopping for a different cost lane, not a better prompt.