Dials on a Curve
Four new anime dials, each on the time curve it is for.
A dial is a strength slider, and the Dial Rack can move that slider during the clip. These four were made with Make Dial from one prompt pair each and tested on the curve shape they are meant for: an impact frame on a 0.5 s pulse, an aura on a ramp, an ink wash as a look, and a colour drain on a snap. Each player runs the baseline and the dial in sync, with the curve and the measured change drawn underneath. Turn the sound on, and use Show to switch to any other strength we rendered.
start hereThe four at a glance
| dial | curve | suggested strengths | verdict |
|---|---|---|---|
| Impact frame | a pulse | pulse 4.5 (text to video) · pulse 6 (with a reference picture) | Ready, on text to video. |
| Aura | ramp and hold | ramp 2.5 (with a reference picture) · ramp 2 (text to video) | Ready, with two notes. |
| Ink wash | a constant look (or seep in and out) | constant 1.5 to 2 (text to video only) · seep 2 at k = 6 | Ready as a constant look on text to video. |
| Colour drain | a snap | snap 4 (with a reference picture); 3 draws nothing | Ready, with a reference picture. |
- The players: the baseline (same prompt and seed, dial off) and the dial's takes play together. Click a video or tick hear to choose which soundtrack you hear. Show lists every take: full-size first, then the low-res tests (640x352 calibration runs, marked as such).
- The chart: the blue area is the curve M(t), from 0 (off) to 1 (the row's full strength, so strength = peak × M). The gold line is the measured change from the baseline, frame by frame, for the take on the right; the dashed line is the same for the other take. If the gold line rises and falls with the blue area, the dial follows its curve. Click the chart to jump.
- Constant means the dial held at full strength for the whole clip. A full-size constant is a time row held at 1 with the same k and audio setting as its curve, so constant vs curve differ only in the curve. A low-res "plain row" is an ordinary rack row on all steps, sound dialled too.
- k (blend steps) is how many of the 6 sampling steps blend the curve frame by frame; k = 6 is all of them. audio: protect keeps the dial off the soundtrack (partly: the sound still hears the changed picture).
- With a reference picture means Pika's reference-to-video with the first frame of the full-size text-to-video baseline as Picture 1 ("The shot opens on <Picture 1>"), the situation a recurring character puts a shot in.
Every clip: Pika H3 Flex, 6 steps (lcm, manual sigmas, cfg 1), 3.75 s at 24 fps, seed 4242, de-rope off. Full-size takes were rendered at 1280x704 and are shown at 960x528. Render time on one RTX 5090: about 1 to 1.5 min a full-size take (k = 6), 15 to 25 s a low-res test. One scene and one seed per dial, so treat the numbers as leads.
a pulseImpact frame
The anime hit flash: for a few frames the shot turns into a stark, inverted black-and-white drawing with heavy ink lines. On a 0.5 s pulse at the moment the punch lands, the two men become black silhouettes on white inside an ink starburst for about six frames, and then the dojo comes back in colour with the same men and the same follow-through. Held over the whole clip, the dial replaces the shot with an ink drawing of one masked figure: the curve is what makes it an impact frame.
| use | strength | what you get |
|---|---|---|
| text to video, pulse | 4.5 | the public example: a full-frame flash, the shot clean before and after |
| with a reference picture, pulse | 6 | an ink burst over the contact rather than a full-frame flash; 8 drew less than 6 |
| text to video, held constant | 3 | turns the whole clip into ink |
Verdict: Ready, on text to video. Curve: step keys [0 s: 0] [1.95 s: 1] [2.45 s: 0]: a 0.5 s pulse starting just before the contact (frame 48 = 2.0 s). Blend steps k = 6, audio: protect.
- The pulse lands on the action, not on the key edges: put the step around the moment of contact.
- Calibrate a pulse at full size. At 640x352 the reference-picture pulse at 6 was a full-frame burst; at 1280x704 it was a partial one.
- With audio: protect the curve take is still ~9.5 LU louder than the baseline (the hit sounds bigger).
The prompt pair that made it: h3_impact_frame.safetensors
# A (−, normal shading) ||| B (+, impact frame)
{subject}, drawn with normal colours and soft shading. ||| {subject}, drawn as a stark black-and-white inverted impact frame with heavy ink lines.
{subject} runs over eight adult subjects: A young woman in a long coat on a stone rooftop; An old man in a grey robe in a library; A knight in plate armour in a castle courtyard; A woman in a forest clearing; A man in a jacket on a city street at night; A traveller in a cloak on a desert dune; A monk on a mountain temple terrace; A woman in an apron in a small kitchen. Pair agreement 0.949, debias on, hook after H3's
text refiner. Download: h3_impact_frame.safetensors ·
example_pairs/impact_frame.txt. The file carries its own description, labels and
strengths (⚙ in the Dial Rack).
ramp and holdAura
The anime power-up: a blazing aura of glowing energy flaring around the body. On a ramp-and-hold curve with a reference picture, nothing changes for the first 1.5 s; then a ring of light flares round her, energy spikes and dust rise, and the camera pushes in to a gritted-teeth close-up that holds to the end. The aura is not named in the prompt: the dial adds it.
| use | strength | what you get |
|---|---|---|
| with a reference picture, ramp | 2.5 | follows the ramp; the opening stays as it was |
| with a reference picture, softer | 2 | a little weaker |
| text to video, ramp | 2 | follows too, but her costume and hair change for the whole clip; 3 bursts from frame 0 |
Verdict: Ready, with two notes. Curve: linear keys [0 s: 0] [0.8 s: 0] [2.2 s: 1] [3.75 s: 1]: off, a 1.4 s ramp, then held. Blend steps k = 6, audio: protect.
- The push-in comes with the dial: it moves the camera in to a close-up, and her face drifts in that close-up (her character score fell from 0.77-0.97 to 0.54-0.80 after the ramp).
- The soundtrack gets ~18 LU louder (a wind roar over a near-silent baseline) even with audio: protect.
- With a reference picture, k = 2 followed the ramp as well as k = 6.
The prompt pair that made it: h3_aura.safetensors
# A (−, no aura) ||| B (+, energy aura)
{subject}, with no glow or light around the body. ||| {subject}, with a blazing aura of glowing energy flaring up around the body.
{subject} runs over eight adult subjects: A young woman in a long coat on a stone rooftop; An old man in a grey robe in a library; A knight in plate armour in a castle courtyard; A woman in a forest clearing; A man in a jacket on a city street at night; A traveller in a cloak on a desert dune; A monk on a mountain temple terrace; A woman in an apron in a small kitchen. Pair agreement 0.928, debias on, hook after H3's
text refiner. Download: h3_aura.safetensors ·
example_pairs/aura.txt. The file carries its own description, labels and
strengths (⚙ in the Dial Rack).
a constant look (or seep in and out)Ink wash
The scene becomes a flowing black ink-wash painting on paper, and the negative side goes to clean cel-shaded anime colour. Its best use is as a constant look on text to video: at 1.5 to 2 the whole walk is a clean, consistent ink-wash painting. On a seep-in-and-out curve it does follow (cel anime, ink by about 1.1 s, colour again by the end), but it plays as a dissolve between two drawings rather than a filter: the pines and peaks re-compose on the way in and out.
| use | strength | what you get |
|---|---|---|
| text to video, constant | 1.5 to 2 | the recommended use: a clean ink-wash clip |
| text to video, seep curve | 2 at k = 6 | follows, but re-composes the landscape at both ends; k = 2 smears the look over the whole clip |
| with a reference picture | none | draws nothing up to 6, curve or constant: the picture holds the style |
Verdict: Ready as a constant look on text to video. Curve: linear keys [0 s: 0] [1.3 s: 1] [2.1 s: 1] [3.4 s: 0] [3.75 s: 0]: in, hold, out. Blend steps k = 6, audio: protect.
- Text to video only. With a reference picture it draws nothing at any strength we tried (2, 4, 6).
- The figure keeps his hat, robe and sword but is redrawn in the ink stretch.
The prompt pair that made it: h3_ink_wash.safetensors
# A (−, cel-shaded colour) ||| B (+, ink wash)
{subject}, drawn in clean cel-shaded anime colours. ||| {subject}, painted as a flowing black ink-wash painting on paper.
{subject} runs over eight adult subjects: A young woman in a long coat on a stone rooftop; An old man in a grey robe in a library; A knight in plate armour in a castle courtyard; A woman in a forest clearing; A man in a jacket on a city street at night; A traveller in a cloak on a desert dune; A monk on a mountain temple terrace; A woman in an apron in a small kitchen. Pair agreement 0.968, debias on, hook after H3's
text refiner. Download: h3_ink_wash.safetensors ·
example_pairs/ink_wash.txt. The file carries its own description, labels and
strengths (⚙ in the Dial Rack).
a snapColour drain
Full rich colour to flat grey monochrome, for a time-stop, a memory or a moment of shock. With a reference picture, the greenhouse stays in colour until the blades meet, drains to flat grey just before the 1.7 s key, holds while the fencers lock, and snaps back to red and green on the 3.0 s key, with the same fencers and moves throughout (a faint trace of the drain stays after).
| use | strength | what you get |
|---|---|---|
| with a reference picture, snap | 4 | full grey on the keys, the same fencers |
| with a reference picture | 3 | draws nothing: there is a threshold |
| text to video | 4 to 5 | greys, but turns the fencers into masked mannequins and restages the set |
Verdict: Ready, with a reference picture. Curve: step keys [0 s: 0] [1.7 s: 1] [3.0 s: 0]: grey as the blades lock, colour back at 3.0 s. Blend steps k = 6, audio: protect.
- Use it with a reference picture (or a first frame you want to keep). Without one it restages the scene even in the colour stretches.
- The drain starts ~0.3 s before the key; put the key a little late.
The prompt pair that made it: h3_colour_drain.safetensors
# A (−, full colour) ||| B (+, grey monochrome)
{subject}, in full, rich natural colour. ||| {subject}, in flat grey monochrome with every colour drained away.
{subject} runs over eight adult subjects: A young woman in a long coat on a stone rooftop; An old man in a grey robe in a library; A knight in plate armour in a castle courtyard; A woman in a forest clearing; A man in a jacket on a city street at night; A traveller in a cloak on a desert dune; A monk on a mountain temple terrace; A woman in an apron in a small kitchen. Pair agreement 0.936, debias on, hook after H3's
text refiner. Download: h3_colour_drain.safetensors ·
example_pairs/colour_drain.txt. The file carries its own description, labels and
strengths (⚙ in the Dial Rack).
notesWhat we learned
- A reference picture holds the style, so a dial's useful strength depends on the pipeline. Ink wash drew nothing with a reference picture up to 6 (1.5 was a full ink painting in text to video); colour drain needed 4 (3 drew nothing); impact frame needed 6 instead of 4.5.
- k = 2 is not always enough. With a reference picture, aura and colour drain followed their curves at k = 2 as well as at k = 6. Without one, the ink-wash curve needed k = 6 (k = 2 smeared the look over the whole clip). Start look dials on curves at k = 6.
- A pulse lands on the action. The 0.5 s impact pulse made a ~6-frame flash right at the contact. More is not always stronger: 8 on the pulse drew less than 6.
- audio: protect is partial when the picture changes a lot: against the baseline, the aura curve take is ~18 LU louder, impact frame ~9.5 LU, ink wash ~5.5 LU; colour drain ~2.5 LU quieter.
- With a reference picture, even a constant row leaves the first ~0.5 s close to the picture (aura, colour drain), so constant and curve differ less at the start there than in text to video.