Visual storytelling reading list (28 August 2026)

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: making a character reference sheet, and what it is actually good for.

• A single photo against a four-view sheet, compared on screen -- a tutorial for a character-consistency LoRA that runs the same scene twice, once conditioned on one portrait and once on a four-column sheet (face close-up, front, 90-degree side, back, white background). Interesting for two findings that do not depend on the generator it happens to use: the sheet holds hairstyle, clothing and body proportions across a scene change better than the single photo does, and a single photo degrades visibly when its aspect ratio does not match the target video's orientation. It also shows its own tiled upscale pass slightly degrading identity, because the second stage redraws the face rather than enlarging it. Watch "Single Photo vs. Character Sheet: The LTX 2.3 Best Face ID Secret" by Veteran AI, published 2026-07-29

• One-shot vs. staged character creation in Qwen Image Edit -- a workflow that composites and re-edits reference characters directly in one pass, contrasted with a separate staged pipeline (turnaround sheet, dataset captioning, LoRA training) built on the same base model. Interesting because the creator states a structural reason to stage beyond output quality: a one-shot edit bakes in the editing model's own rendering style, so the result can't be re-rendered by a different downstream video model. Watch "Create HYPERREALISTIC Consistent AI Characters - FREE & LOCAL! [Full ComfyUI Masterclass 2025]" by Mickmumpitz, published 2025-10-07

• A local pipeline that scores its own consistency -- a two-week build that generates a three-panel sheet in one seed, crops it into face, half-body and full-body anchor tiles, and drives batches of angle and lifestyle generations from those anchors across three models, auto-sorting every output by cosine face-distance from the reference. Interesting for the scoring step, which puts a number on something usually judged by eye, and for the on-screen comparison behind it: generations referencing the full three-panel grid score closer to the reference than ones referencing a single portrait. The video is explicit that curation stays manual and that not every output is usable. Watch "Free Local AI Character Dataset Generator: Krea 2, Qwen Edit & Flux 2 Klein (ComfyUI Deep Dive)" by The 3-Minute Node, published 2026-07-27

• A prompt-driven sheet with the prompts published -- generating a reference sheet from one uploaded photo in Nano Banana: four full-body views on the top row, three close-up portraits on the bottom, then feeding the finished sheet back in as a reference alongside a start frame. Interesting because the three prompts are given in full in the description (with reference image, without one, and a wardrobe-change variant), and because the working prompt runs to roughly 180 words of explicit instruction covering pose, spacing, lighting direction and head-height uniformity. The video shows two generated sheets and one picked by eye, not a repeat test across seeds. Watch "Consistent Character Sheets in Nano Banana (Prompts Included!)" by AI Video School, published 2026-01-23

• Batch the angles, then spot-fix the bad one -- a short walkthrough of generating many camera angles of one character from a single reference photo, using Qwen-Image-Edit-2511 with a Lightning LoRA, a separate multi-angle LoRA and a third-party node giving a 3D azimuth/elevation/zoom viewport. Interesting for the working method it states out loud: generate the batch, then use a single-angle override node to redo only the one shot that came out wrong, rather than re-running everything. Worth knowing before clicking that the title's "360 turnarounds" means a batch of sampled angles aimed at a 20-30 image LoRA dataset -- no continuous full-circle sweep is shown -- and that the multi-angle behaviour comes from the LoRA and custom node, not from the base model. Watch "Qwen Image Edit 2511: Create 360-degree Character Turnarounds in ComfyUI" by smallzero, published 2026-02-14

• Krea's own description of what Krea 2 is for -- a vendor blog post naming four character reference-sheet types the model targets (turnarounds, expression sheets, costume variations, key portraits), with worked examples and a published prompt template. Interesting for its own stated limitation: turnarounds "will be visually consistent -- same outfit, same hair color, same proportions -- but not pixel-perfect." Read "Character Design with Krea 2" by The Krea Team, published 2026-05-23