Repository navigation
Examples
Every clip on this page was generated inside Forge Neo by this extension, from text, with the pan2 preset's recipe. Nothing was upscaled, retimed or edited afterwards: these are the MP4 files as Forge saved them. PAN 2 makes video only, so the clips are silent.
Click a preview to download the MP4. The previews are GIFs at a lower frame rate and size.
Common setup: one NVIDIA RTX 6000 Ada (48 GB) with 62 GB of system RAM on RunPod, Forge Neo neo 2.32 (a31a8ba0), 24 FPS. Model files: PAN2_int8_convrot_full, PAN2_qwen3_5_int8_convrot and PAN2_vae_fp16. Euler, Simple, 50 steps, CFG 2.5, Shift 9, seed 123, and the negative prompt left empty (the released one is used). Times are the full request through Forge's API, with the model already loaded unless noted.
Note
These clips were made with the code of v0.1.0.
- A story with camera moves: skatepark, 8 seconds
- Built on the Regional Boxes timeline: busker and dancer, 5 seconds
- PAN 2 and MiniMax H3, same scene
- Trained size: potter, 5 seconds
- Writing the description
Four events and six camera segments in one continuous shot: the skater drops in (static, then tilt_down + dolly_forward), carves around the bowl with the dog running along the rim (track_right + pan_right), launches a grabbed 180 above the coping (tilt_up + dolly_up, then tilt_down + unstable), and rolls out with her arms up while the view rises and pulls back over the park and the city (dolly_backward + dolly_up).
The dog leaves the picture during the air, where the description gives it no action between 4.5 and 6.5 seconds, and comes back at the end as the same dog, blue collar and all: PAN 2 keeps track of it while it is out of view. To keep a subject on screen the whole time, give it an action in every event.
| Model | Size | Frames | Steps | CFG | Shift | Seed | Time |
|---|---|---|---|---|---|---|---|
| INT8 | 832×480 | 193 (8 s) | 50 | 2.5 | 9 | 123 | 844.6 s (13.7 s/step), model switched in |
Description
{"video_style":"Realistic action sports cinematography, handheld follow camera, in one uninterrupted continuous shot.","lighting":"Low golden-hour sun from the right, long shadows across pale concrete, warm rim light on the riders.","perspective_view":"Starts at eye level on the coping of a deep concrete bowl, then follows the skater through the bowl at close range and finishes high above the park.","subjects_lifecycle":[{"subject_id":"<|Subject1|>","display_name":"skateboarder","subject_attribute":"a young woman in a red hoodie, black ripped jeans and white sneakers, curly hair tied back, riding a worn maple skateboard with orange wheels"},{"subject_id":"<|Subject2|>","display_name":"dog","subject_attribute":"a lean brown and white border collie with a blue collar"}],"init_frame_visual_description":"A skateboarder stands on the coping at the edge of a deep, empty concrete skate bowl, the tail of her board pinned under her back foot with the nose hanging over the drop. A border collie sits on the flat deck beside her, ears up. Graffiti covers the far wall of the bowl, and palm trees and a city skyline sit low on the horizon.","events":[{"event_index":0,"time_range_sec":"0 to 2","event_description":"The skater leans forward, stomps the nose down and drops into the bowl, accelerating down the steep transition while the dog jumps up and starts running along the rim.","camera_motion":[{"time_range_sec":"0 to 0.4","camera_motion":["static"]},{"time_range_sec":"0.4 to 2","camera_motion":["tilt_down","dolly_forward"]}],"contained_actions":[{"time_range_sec":"0.3 to 1","action":"<|Subject1|> shifts her weight onto her front foot and drops the board into the bowl"},{"time_range_sec":"1 to 2","action":"<|Subject1|> rolls down the transition and crouches low, picking up speed across the flat bottom"},{"time_range_sec":"0.6 to 2","action":"<|Subject2|> springs up and sprints along the rim of the bowl"}]},{"event_index":1,"time_range_sec":"2 to 4.5","event_description":"She carves a wide arc across the bowl, sneakers scraping the deck and hand trailing along the wall, while the dog races level with her along the rim above.","camera_motion":[{"time_range_sec":"2 to 4.5","camera_motion":["track_right","pan_right"]}],"contained_actions":[{"time_range_sec":"2 to 4.5","action":"<|Subject1|> carves a fast frontside arc around the bowl wall, trailing her left hand on the concrete"},{"time_range_sec":"2 to 4.5","action":"<|Subject2|> runs along the rim, keeping pace with the skater"}]},{"event_index":2,"time_range_sec":"4.5 to 6.5","event_description":"Racing up the far wall, she launches above the coping into the air, grabs the edge of the board, turns 180 degrees and drops back down onto the wall, landing with both feet on the bolts.","camera_motion":[{"time_range_sec":"4.5 to 5.5","camera_motion":["tilt_up","dolly_up"]},{"time_range_sec":"5.5 to 6.5","camera_motion":["tilt_down","unstable"]}],"contained_actions":[{"time_range_sec":"4.5 to 5.2","action":"<|Subject1|> pumps up the far wall and launches above the coping"},{"time_range_sec":"5.2 to 5.8","action":"<|Subject1|> grabs the board with her front hand and rotates 180 degrees in the air"},{"time_range_sec":"5.8 to 6.5","action":"<|Subject1|> lands back on the wall and rides down into the bowl"}]},{"event_index":3,"time_range_sec":"6.5 to 8","event_description":"She rolls out across the flat bottom, raising both arms, as the dog leaps down into the bowl and runs to meet her. The view rises and pulls back to show the whole skatepark and the city behind it.","camera_motion":[{"time_range_sec":"6.5 to 8","camera_motion":["dolly_backward","dolly_up"]}],"contained_actions":[{"time_range_sec":"6.5 to 8","action":"<|Subject1|> rolls across the flat bottom with both arms raised"},{"time_range_sec":"6.8 to 8","action":"<|Subject2|> jumps down into the bowl and runs to the skater"}]}],"quality":{"video_quality":"Sharp, clean detail on the skater and dog with natural motion blur on fast movement, stable exposure through the move.","text_logos_overlays":null,"artifacts":null}}This description was not typed: it was built in the Regional Boxes panel, which turns into a timeline of the clip with the pan2 preset. The look, the light and the first frame go in plain fields, the two subjects get a name and a look, the clip is split into three events by dragging the lines between them, each event gets a camera move from the chips (static, track_right, dolly_backward), and each subject's actions are drawn by dragging on its lane. Write to prompt wrote the JSON below into the prompt box, and Generate made the clip.
Both subjects have an action in every event, so both stay in view from the first beat to the last.
| Model | Size | Frames | Steps | CFG | Shift | Seed | Time |
|---|---|---|---|---|---|---|---|
| INT8 | 832×512 | 121 (5 s) | 50 | 2.5 | 9 | 123 | about 8 s/step, from the txt2img tab |
Description, as Regional Boxes wrote it
{"video_style":"realistic handheld street footage, one continuous shot","lighting":"warm evening light from neon shop signs and string lights, wet reflective pavement after rain","perspective_view":"eye-level medium-wide shot","subjects_lifecycle":[{"subject_id":"<|Subject1|>","display_name":"busker","subject_attribute":"a young man with dreadlocks in a yellow raincoat, drumming on upturned plastic buckets"},{"subject_id":"<|Subject2|>","display_name":"dancer","subject_attribute":"a woman in a red hoodie, black leggings and white sneakers"}],"init_frame_visual_description":"A narrow market street at dusk after rain, the wet cobblestones reflecting pink and blue neon signs. A busker in a yellow raincoat sits on a wooden crate behind three upturned plastic buckets, drumsticks in his hands. A woman in a red hoodie stands next to him on the cobblestones. Behind them a few people shelter under shop awnings.","events":[{"event_index":0,"time_range_sec":"0 to 1.5","event_description":"The busker starts a fast beat on the buckets and the woman starts moving to it.","camera_motion":[{"time_range_sec":"0 to 1.5","camera_motion":["static"]}],"contained_actions":[{"time_range_sec":"0 to 1.5","action":"<|Subject1|> hits the buckets with a fast beat"},{"time_range_sec":"0 to 1.5","action":"<|Subject2|> nods to the beat and steps forward onto the open cobblestones"}]},{"event_index":1,"time_range_sec":"1.5 to 3.5","event_description":"The woman dances across the wet street, spinning and splashing the puddles, while the busker keeps drumming.","camera_motion":[{"time_range_sec":"1.5 to 3.5","camera_motion":["track_right"]}],"contained_actions":[{"time_range_sec":"1.5 to 3.5","action":"<|Subject2|> spins twice on the wet cobblestones, splashing water from the puddles"},{"time_range_sec":"1.5 to 3.5","action":"<|Subject1|> keeps drumming and looks up at her, grinning"}]},{"event_index":2,"time_range_sec":"3.5 to 5","event_description":"She ends with a stomp and both arms up as the busker hits a last beat, and the people under the awnings clap.","camera_motion":[{"time_range_sec":"3.5 to 5","camera_motion":["dolly_backward"]}],"contained_actions":[{"time_range_sec":"3.5 to 5","action":"<|Subject1|> hits a last loud beat and raises both drumsticks"},{"time_range_sec":"3.5 to 5","action":"<|Subject2|> stomps her foot and throws both arms up"}]}],"quality":{"video_quality":"","text_logos_overlays":null,"artifacts":null}}The Puppy, 5 s example from the MiniMax H3 extension (left, with its sound) next to PAN 2 (right) at the same size and length. H3's prompt was rewritten as a PAN 2 description, the way PAN 2's own prompt rewriter does: the puppy and the two children as subjects, three timed events (a leap at a bubble, a spin, a second leap) and a low tracking camera.
H3 moves more like a handheld camera and makes the sound; PAN 2 follows the description closely - the red collar, the girl's yellow t-shirt and wand, the boy's striped shirt and bottle, the spin. The two runs are not a speed comparison: H3 ran on an RTX 4090 at 20 steps and CFG 1, PAN 2 on an RTX 6000 Ada at 50 steps and CFG 2.5 (two model passes per step). PAN 2's description also says more than H3's prompt. PAN 2 was trained at 1248×704, so 640×384 is well below its best.
| Model | Size | Frames | Steps | CFG | Time |
|---|---|---|---|---|---|
| MiniMax H3, INT8 + INT8 text encoder (details) | 640×384 | 124 (5.2 s) | 20 | 1 | 116 s (RTX 4090) |
| PAN 2, INT8 | 640×384 | 121 (5 s) | 50 | 2.5 | 281.6 s (RTX 6000 Ada) |
PAN 2's clip alone: puppy-5s.mp4.
MiniMax H3 prompt
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a low tracking shot at a puppy's height frames a golden retriever puppy on a sunny backyard lawn as soap bubbles drift past. The camera tracks the puppy at fast speed as it leaps, snaps at the bubbles and spins around, while two children in the background blow more bubbles and laugh.
overall_soundscape: The children laugh and the puppy barks excitedly over a light breeze and birdsong. Its paws thump and rustle across the grass.
non_diegetic_music: N/A
PAN 2 description
{"video_style":"Live-action cinematic footage, low handheld tracking camera at a puppy's height, in one uninterrupted continuous shot.","lighting":"Bright midday sun on a backyard lawn, crisp shadows under the puppy, soap bubbles catching rainbow highlights.","perspective_view":"Low tracking view a few centimetres above the grass, close to the puppy, with the backyard and the children softly out of focus behind it.","subjects_lifecycle":[{"subject_id":"<|Subject1|>","display_name":"puppy","subject_attribute":"a fluffy golden retriever puppy with floppy ears and a red collar"},{"subject_id":"<|Subject2|>","display_name":"girl","subject_attribute":"a young girl in a yellow t-shirt holding a bubble wand"},{"subject_id":"<|Subject3|>","display_name":"boy","subject_attribute":"a young boy in a striped shirt holding a bottle of bubble soap"}],"init_frame_visual_description":"A golden retriever puppy crouches on a sunny green lawn, front paws stretched out and tail up, watching a cluster of soap bubbles drifting in front of its nose. Behind it, two children stand on the grass near a wooden fence, the girl raising a bubble wand to her mouth.","events":[{"event_index":0,"time_range_sec":"0 to 1.8","event_description":"The puppy springs up off the grass at the nearest bubble and snaps at it, the bubble bursting against its nose, while the girl blows a new stream of bubbles that drift toward the lawn.","camera_motion":[{"time_range_sec":"0 to 1.8","camera_motion":["track_right"]}],"contained_actions":[{"time_range_sec":"0 to 0.8","action":"<|Subject1|> pushes off with its hind legs and leaps toward a bubble"},{"time_range_sec":"0.8 to 1.8","action":"<|Subject1|> snaps its jaws at the bubble and lands on its front paws"},{"time_range_sec":"0.3 to 1.8","action":"<|Subject2|> blows through the wand, sending a stream of bubbles across the yard"}]},{"event_index":1,"time_range_sec":"1.8 to 3.4","event_description":"The puppy spins around in a full circle chasing a bubble that floats behind it, ears flapping, as the boy laughs and dips another wand into the bottle.","camera_motion":[{"time_range_sec":"1.8 to 3.4","camera_motion":["track_right","pan_left"]}],"contained_actions":[{"time_range_sec":"1.8 to 3.4","action":"<|Subject1|> spins around in a tight circle, chasing a bubble behind it"},{"time_range_sec":"1.8 to 3.4","action":"<|Subject3|> laughs and dips a wand into the bottle of bubble soap"}]},{"event_index":2,"time_range_sec":"3.4 to 5","event_description":"The puppy leaps again into the new cloud of bubbles, popping two of them in a row, and bounds forward across the lawn while both children laugh.","camera_motion":[{"time_range_sec":"3.4 to 5","camera_motion":["track_right"]}],"contained_actions":[{"time_range_sec":"3.4 to 4.3","action":"<|Subject1|> jumps into the cloud of bubbles and snaps two of them"},{"time_range_sec":"4.3 to 5","action":"<|Subject1|> bounds forward across the grass"},{"time_range_sec":"3.4 to 5","action":"<|Subject2|> laughs and waves the wand, releasing more bubbles"},{"time_range_sec":"3.4 to 5","action":"<|Subject3|> laughs and points at the puppy"}]}],"quality":{"video_quality":"Sharp detail on the puppy's fur and the thin bubble films, natural motion blur on fast leaps.","text_logos_overlays":null,"artifacts":null}}The short example from the README at PAN 2's trained size. One event and a static camera: the clay grows into a tall cylinder over the five seconds while the face, the shelves and the light stay steady.
| Model | Size | Frames | Steps | CFG | Shift | Seed | Time |
|---|---|---|---|---|---|---|---|
| INT8 | 1248×704 | 121 (5 s) | 50 | 2.5 | 9 | 123 | 912.2 s (16.8 s/step) |
At this size the RTX 6000 Ada used 46 of its 48 GB: longer clips at 1248×704 need an 80-96 GB card.
Description
{"video_style":"realistic handheld footage, one continuous shot","lighting":"soft overcast daylight","perspective_view":"eye-level medium shot","subjects_lifecycle":[{"subject_id":"<|Subject1|>","display_name":"potter","subject_attribute":"a woman in a clay-stained apron"}],"init_frame_visual_description":"A potter sits at a spinning wheel with a wet lump of clay, shelves of bowls behind her.","events":[{"event_index":0,"time_range_sec":"0 to 5","event_description":"She presses her thumbs into the clay and draws it up into a tall cylinder.","camera_motion":[{"time_range_sec":"0 to 5","camera_motion":["static"]}],"contained_actions":[{"time_range_sec":"0 to 5","action":"<|Subject1|> shapes the clay into a cylinder"}]}],"quality":{"video_quality":"sharp natural detail","text_logos_overlays":null,"artifacts":null}}PAN 2 was trained on structured descriptions with seven keys, in this order: video_style, lighting, perspective_view, subjects_lifecycle, init_frame_visual_description, events and quality. IFM's prompt rewriter skill turns a short idea into one with any chat assistant. What made the difference in these tests:
- One shot. The whole clip is one uninterrupted take; events are phases of it, never cuts.
-
Camera tags come from a fixed list:
static,pan_left,pan_right,tilt_up,tilt_down,track_left,track_right,dolly_forward,dolly_backward,dolly_up,dolly_down,zoom_in,zoom_out,roll_clockwise,roll_counterclockwise,unstable. Combine motions that go together in one segment (track_right+pan_right);staticgoes alone. -
Subject IDs (
<|Subject1|>...) appear only insubjects_lifecycleand at the start of each action; the descriptions use plain words. - Give every subject an action for the whole clip, or it may leave the picture (see the dog above).
- Times cover the clip from 0 to its length with no gaps; 121 frames are 5 seconds, 193 are 8, 241 are 10.



