jbilcke-hf's picture
jbilcke-hf HF staff
add a login wall
b48537f
raw
history blame
No virus
4.68 kB
import { predict } from "./predict"
import { Preset } from "../engine/presets"
import { GeneratedPanel } from "@/types"
import { cleanJson } from "@/lib/cleanJson"
import { createZephyrPrompt } from "@/lib/createZephyrPrompt"
import { dirtyGeneratedPanelCleaner } from "@/lib/dirtyGeneratedPanelCleaner"
import { dirtyGeneratedPanelsParser } from "@/lib/dirtyGeneratedPanelsParser"
import { sleep } from "@/lib/sleep"
export const predictNextPanels = async ({
preset,
prompt = "",
nbPanelsToGenerate = 2,
nbTotalPanels = 4,
existingPanels = [],
}: {
preset: Preset;
prompt: string;
nbPanelsToGenerate?: number;
nbTotalPanels?: number;
existingPanels: GeneratedPanel[];
}): Promise<GeneratedPanel[]> => {
// console.log("predictNextPanels: ", { prompt, nbPanelsToGenerate })
// throw new Error("Planned maintenance")
// In case you need to quickly debug the RENDERING engine you can uncomment this:
// return mockGeneratedPanels
const existingPanelsTemplate = existingPanels.length
? ` To help you, here are the previous panels and their captions (note: if you see an anomaly here eg. no caption or the same description repeated multiple times, do not hesitate to fix the story): ${JSON.stringify(existingPanels, null, 2)}`
: ''
const firstNextOrLast =
existingPanels.length === 0
? "first"
: (nbTotalPanels - existingPanels.length) === nbTotalPanels
? "last"
: "next"
const query = createZephyrPrompt([
{
role: "system",
content: [
`You are a writer specialized in ${preset.llmPrompt}`,
`Please write detailed drawing instructions and short (2-3 sentences long) speech captions for the ${firstNextOrLast} ${nbPanelsToGenerate} panels (out of ${nbTotalPanels} in total) of a new story, but keep it open-ended (it will be continued and expanded later). Please make sure each of those ${nbPanelsToGenerate} panels include info about character gender, age, origin, clothes, colors, location, lights, etc. Only generate those ${nbPanelsToGenerate} panels, but take into account the fact the panels are part of a longer story (${nbTotalPanels} panels long).`,
`Give your response as a VALID JSON array like this: \`Array<{ panel: number; instructions: string; caption: string; }>\`.`,
// `Give your response as Markdown bullet points.`,
`Be brief in the instructions and narrative captions of those ${nbPanelsToGenerate} panels, don't add your own comments. The captions must be captivating, smart, entertaining. Be straight to the point, and never reply things like "Sure, I can.." etc. Reply using valid JSON!! Important: Write valid JSON!`
].filter(item => item).join("\n")
},
{
role: "user",
content: `The story is about: ${prompt}.${existingPanelsTemplate}`,
}
]) + "\n[{"
let result = ""
// we don't require a lot of token for our task
// but to be safe, let's count ~130 tokens per panel
const nbTokensPerPanel = 130
const nbMaxNewTokens = nbPanelsToGenerate * nbTokensPerPanel
try {
// console.log(`calling predict(${query}, ${nbTotalPanels})`)
result = `${await predict(query, nbMaxNewTokens)}`.trim()
console.log("LLM result (1st trial):", result)
if (!result.length) {
throw new Error("empty result on 1st trial!")
}
} catch (err) {
// console.log(`prediction of the story failed, trying again..`)
// this should help throttle things on a bit on the LLM API side
await sleep(2000)
try {
result = `${await predict(query + " \n ", nbMaxNewTokens)}`.trim()
console.log("LLM result (2nd trial):", result)
if (!result.length) {
throw new Error("empty result on 2nd trial!")
}
} catch (err) {
console.error(`prediction of the story failed twice 💩`)
throw new Error(`failed to generate the story twice 💩 ${err}`)
}
}
// console.log("Raw response from LLM:", result)
const tmp = cleanJson(result)
let generatedPanels: GeneratedPanel[] = []
try {
generatedPanels = dirtyGeneratedPanelsParser(tmp)
} catch (err) {
// console.log(`failed to read LLM response: ${err}`)
// console.log(`original response was:`, result)
// in case of failure here, it might be because the LLM hallucinated a completely different response,
// such as markdown. There is no real solution.. but we can try a fallback:
generatedPanels = (
tmp.split("*")
.map(item => item.trim())
.map((cap, i) => ({
panel: i,
caption: cap,
instructions: cap,
}))
)
}
return generatedPanels.map(res => dirtyGeneratedPanelCleaner(res))
}