Visual storytelling reading list (29 August 2026)

August 29, 2026By Alan Kent · AI agent architect; building Ordinary Animator

This list is generated by an AI research process and is explicitly not editorial recommendation. Interest is not endorsement.

Everything below is about one thing: getting a subject, or a camera, to end up where you meant it to be.

Copy a camera move off a rough 3D render -- a workflow that feeds a crude Blender driving video plus an aligned first frame into a camera-transfer LoRA on LTX 2.3, so the generated shot inherits the render's camera movement rather than being described in words. Interesting for the two failure cases it shows rather than claims: a subject the model cannot confidently classify drifts identity mid-turn (a wolf goes horse-like) until an explicit subject prompt is added, and a scene with obstacle-crossing and complex character action fails across several regenerations, not one bad seed. Also states a concrete reason to leave the prompt blank by default, since an action prompt competes with the camera signal coming from the driving video. Watch "Master Pro Camera Control: Transfer 3D Motions to LTX 2.3" by Veteran AI, published 2026-07-06

Make the driving video when you do not have one -- the same control problem approached from the other end: reconstruct a 3D scene or a rigged 3D body from a single still using Meta's SAM 3D playground, orbit it by hand, screen-record that navigation, and feed the recording into Uni3C as the camera path for Wan 2.1. Interesting because it removes the usual precondition for reference-driven camera control, which is having an existing clip to copy from. Worth knowing before starting: the creator's own side-by-side finds body-model guidance noticeably stronger than scene-model guidance, and the whole route is several more steps than a simple pan or orbit preset, with no comparison against one shown. Watch "Uni3C + SAM 3D: The Ultimate Workflow for AI Camera Control" by Veteran AI, published 2025-11-21

A mouse-drawn sketch as the composition input -- a rough sketch drawn in MS Paint, deliberately, to prove no drawing tablet is needed, run through a lineart ControlNet to fix the composition of a shot, with re-prompting changing setting and time of day while the layout holds. For two characters, the sketch gets colour-coded regions that become masks, so each character gets its own prompt. Interesting for a before-and-after that is on screen rather than asserted: a single whole-image prompt visibly merges the two characters, and switching to colour regions fixes it. Two honest caveats from the creator, both stated on camera: pushing the composition to an extreme angle was simply not honoured, and results are not first-try. Worth knowing that this is a mid-2024 stack, so treat the pattern as the transferable part and the specific models as history. Watch "Next level AI Filmmaking: Turn your storyboard sketches into MOVIES!" by Mickmumpitz, published 2024-07-19

Block the scene in 3D, then mostly turn the depth map off -- a multi-character pipeline that builds a Blender layout to lock proportions, pose and camera before any generation, with per-character LoRAs and regional hooks to stop two characters bleeding into each other. Interesting for a candid detail that cuts against how depth control is usually pitched: the creator says he deactivates the depth ControlNet about 95 percent of the time, using the 3D layout as a blocking aid rather than as a signal fed to the model. The video ends on a finished short, so the pipeline is shown carrying through rather than stopping at stills. Watch "Control MULTIPLE CONSISTENT CHARACTERS + CAMERA with this FREE AI Workflow (Blender + ComfyUI)" by Mickmumpitz, published 2025-03-07

A pose LoRA that says plainly what it is not -- an OpenPose-style control LoRA for Krea 2 Turbo, from an independent developer rather than Krea's own team, which follows body keypoints while the prompt defines identity, clothing, style and background. Interesting for one sentence in its own model card: it describes itself as a straightforward pose control LoRA and explicitly not a reference-plus-pose fusion model. That is the author ruling out, in their own words, the combination most people actually want, which is holding a pose and keeping a specific character's face from a reference image at the same time. Read "Krea-2-pose-controlnet" by thedeoxen

Visual storytelling reading list (29 August 2026)

August 29, 2026

By Alan Kent · AI agent architect; building Ordinary Animator

This list is generated by an AI research process and is explicitly not editorial recommendation. Interest is not endorsement.

Everything below is about one thing: getting a subject, or a camera, to end up where you meant it to be.

Copy a camera move off a rough 3D render -- a workflow that feeds a crude Blender driving video plus an aligned first frame into a camera-transfer LoRA on LTX 2.3, so the generated shot inherits the render's camera movement rather than being described in words. Interesting for the two failure cases it shows rather than claims: a subject the model cannot confidently classify drifts identity mid-turn (a wolf goes horse-like) until an explicit subject prompt is added, and a scene with obstacle-crossing and complex character action fails across several regenerations, not one bad seed. Also states a concrete reason to leave the prompt blank by default, since an action prompt competes with the camera signal coming from the driving video. Watch "Master Pro Camera Control: Transfer 3D Motions to LTX 2.3" by Veteran AI, published 2026-07-06

Make the driving video when you do not have one -- the same control problem approached from the other end: reconstruct a 3D scene or a rigged 3D body from a single still using Meta's SAM 3D playground, orbit it by hand, screen-record that navigation, and feed the recording into Uni3C as the camera path for Wan 2.1. Interesting because it removes the usual precondition for reference-driven camera control, which is having an existing clip to copy from. Worth knowing before starting: the creator's own side-by-side finds body-model guidance noticeably stronger than scene-model guidance, and the whole route is several more steps than a simple pan or orbit preset, with no comparison against one shown. Watch "Uni3C + SAM 3D: The Ultimate Workflow for AI Camera Control" by Veteran AI, published 2025-11-21

A mouse-drawn sketch as the composition input -- a rough sketch drawn in MS Paint, deliberately, to prove no drawing tablet is needed, run through a lineart ControlNet to fix the composition of a shot, with re-prompting changing setting and time of day while the layout holds. For two characters, the sketch gets colour-coded regions that become masks, so each character gets its own prompt. Interesting for a before-and-after that is on screen rather than asserted: a single whole-image prompt visibly merges the two characters, and switching to colour regions fixes it. Two honest caveats from the creator, both stated on camera: pushing the composition to an extreme angle was simply not honoured, and results are not first-try. Worth knowing that this is a mid-2024 stack, so treat the pattern as the transferable part and the specific models as history. Watch "Next level AI Filmmaking: Turn your storyboard sketches into MOVIES!" by Mickmumpitz, published 2024-07-19

Block the scene in 3D, then mostly turn the depth map off -- a multi-character pipeline that builds a Blender layout to lock proportions, pose and camera before any generation, with per-character LoRAs and regional hooks to stop two characters bleeding into each other. Interesting for a candid detail that cuts against how depth control is usually pitched: the creator says he deactivates the depth ControlNet about 95 percent of the time, using the 3D layout as a blocking aid rather than as a signal fed to the model. The video ends on a finished short, so the pipeline is shown carrying through rather than stopping at stills. Watch "Control MULTIPLE CONSISTENT CHARACTERS + CAMERA with this FREE AI Workflow (Blender + ComfyUI)" by Mickmumpitz, published 2025-03-07

A pose LoRA that says plainly what it is not -- an OpenPose-style control LoRA for Krea 2 Turbo, from an independent developer rather than Krea's own team, which follows body keypoints while the prompt defines identity, clothing, style and background. Interesting for one sentence in its own model card: it describes itself as a straightforward pose control LoRA and explicitly not a reference-plus-pose fusion model. That is the author ruling out, in their own words, the combination most people actually want, which is holding a pose and keeping a specific character's face from a reference image at the same time. Read "Krea-2-pose-controlnet" by thedeoxen