forked from cgvr/DeltaVR
543 lines
19 KiB
C#
543 lines
19 KiB
C#
|
|
using System;
|
|
using System.Diagnostics;
|
|
using System.Net.Http;
|
|
using System.Security.Cryptography;
|
|
using System.Text;
|
|
using System.Threading.Tasks;
|
|
using UnityEngine;
|
|
using Valve.Newtonsoft.Json;
|
|
using Valve.Newtonsoft.Json.Linq;
|
|
|
|
|
|
public class InvokeAiClient : MonoBehaviour
|
|
{
|
|
public static InvokeAiClient Instance { get; private set; }
|
|
|
|
public string INVOKEAI_BASE_URL;
|
|
public string DEFAULT_QUEUE_ID = "default";
|
|
public string MODEL_KEY;
|
|
|
|
private HttpClient httpClient;
|
|
|
|
private void Awake()
|
|
{
|
|
httpClient = new HttpClient
|
|
{
|
|
Timeout = TimeSpan.FromSeconds(120)
|
|
};
|
|
httpClient.BaseAddress = new Uri(INVOKEAI_BASE_URL);
|
|
|
|
Instance = this;
|
|
}
|
|
|
|
// Start is called before the first frame update
|
|
void Start()
|
|
{
|
|
|
|
}
|
|
|
|
// Update is called once per frame
|
|
void Update()
|
|
{
|
|
|
|
}
|
|
|
|
private async Task<JArray> ListModels(string modelType = "main")
|
|
{
|
|
var requestUri = $"/api/v2/models/?model_type={Uri.EscapeDataString(modelType)}";
|
|
using var resp = await httpClient.GetAsync(requestUri).ConfigureAwait(false);
|
|
resp.EnsureSuccessStatusCode();
|
|
|
|
var json = await resp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
var root = JObject.Parse(json);
|
|
return (JArray) root["models"];
|
|
}
|
|
|
|
private async Task<JObject> GetModelInfo(string modelKey)
|
|
{
|
|
var requestUri = $"/api/v2/models/i/{Uri.EscapeDataString(modelKey)}";
|
|
using var resp = await httpClient.GetAsync(requestUri).ConfigureAwait(false);
|
|
resp.EnsureSuccessStatusCode();
|
|
|
|
var json = await resp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
return JObject.Parse(json);
|
|
}
|
|
|
|
private async Task<string> GetImageUrl(string imageName)
|
|
{
|
|
var requestUri = $"/api/v1/images/i/{Uri.EscapeDataString(imageName)}/urls";
|
|
UnityEngine.Debug.Log("Get image URL: " + requestUri);
|
|
using var resp = await httpClient.GetAsync(requestUri).ConfigureAwait(false);
|
|
resp.EnsureSuccessStatusCode();
|
|
|
|
var json = await resp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
var root = JObject.Parse(json);
|
|
return root.Value<string>("image_url");
|
|
}
|
|
|
|
|
|
private async Task<JObject> WaitForCompletion(string batchId, int timeoutSeconds = 300)
|
|
{
|
|
var sw = Stopwatch.StartNew();
|
|
string queueId = DEFAULT_QUEUE_ID;
|
|
|
|
while (true)
|
|
{
|
|
if (sw.Elapsed.TotalSeconds > timeoutSeconds)
|
|
throw new TimeoutException($"Image generation timed out after {timeoutSeconds} seconds");
|
|
|
|
// Get batch status
|
|
var statusUrl = $"/api/v1/queue/{Uri.EscapeDataString(queueId)}/b/{Uri.EscapeDataString(batchId)}/status";
|
|
using var statusResp = await httpClient.GetAsync(statusUrl).ConfigureAwait(false);
|
|
statusResp.EnsureSuccessStatusCode();
|
|
|
|
var statusJson = await statusResp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
var statusData = JObject.Parse(statusJson);
|
|
|
|
// Check for failures
|
|
int failedCount = statusData.Value<int?>("failed") ?? 0;
|
|
if (failedCount > 0)
|
|
{
|
|
var queueStatusUrl = $"/api/v1/queue/{Uri.EscapeDataString(queueId)}/status";
|
|
using var queueResp = await httpClient.GetAsync(queueStatusUrl).ConfigureAwait(false);
|
|
queueResp.EnsureSuccessStatusCode();
|
|
|
|
var queueJson = await queueResp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
var queueData = JObject.Parse(queueJson);
|
|
|
|
throw new InvalidOperationException(
|
|
$"Image generation failed. Batch {batchId} has {failedCount} failed item(s). " +
|
|
$"Queue status: {queueData.ToString(Formatting.Indented)}"
|
|
);
|
|
}
|
|
|
|
// Check completion
|
|
int completed = statusData.Value<int?>("completed") ?? 0;
|
|
int total = statusData.Value<int?>("total") ?? 0;
|
|
|
|
if (completed == total && total > 0)
|
|
{
|
|
// Get most recent non-intermediate image
|
|
const string imagesPath = "/api/v1/images/?is_intermediate=false&limit=10";
|
|
using var imagesResp = await httpClient.GetAsync(imagesPath).ConfigureAwait(false);
|
|
imagesResp.EnsureSuccessStatusCode();
|
|
|
|
var imagesJson = await imagesResp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
var imagesData = JObject.Parse(imagesJson);
|
|
|
|
var items = imagesData["items"] as JArray;
|
|
if (items != null && items.Count > 0)
|
|
{
|
|
var imageName = items[0].Value<string>("image_name");
|
|
|
|
// Return result object mirroring your Python structure
|
|
var result = new JObject
|
|
{
|
|
["batch_id"] = batchId,
|
|
["status"] = "completed",
|
|
["result"] = new JObject
|
|
{
|
|
["outputs"] = new JObject
|
|
{
|
|
["save_image"] = new JObject
|
|
{
|
|
["type"] = "image_output",
|
|
["image"] = new JObject
|
|
{
|
|
["image_name"] = imageName
|
|
}
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
return result;
|
|
}
|
|
|
|
// If no images found, return the status object
|
|
return statusData;
|
|
}
|
|
|
|
// Wait before checking again
|
|
await Task.Delay(1000).ConfigureAwait(false);
|
|
}
|
|
}
|
|
|
|
|
|
private async Task<JObject> EnqueueGraph(JToken graph)
|
|
{
|
|
string queueId = DEFAULT_QUEUE_ID;
|
|
|
|
// Build request JSON dynamically
|
|
var payload = new JObject
|
|
{
|
|
["batch"] = new JObject
|
|
{
|
|
["graph"] = graph, // graph can be any JSON structure
|
|
["runs"] = 1,
|
|
["data"] = JValue.CreateNull()
|
|
}
|
|
};
|
|
|
|
var url = $"/api/v1/queue/{Uri.EscapeDataString(queueId)}/enqueue_batch";
|
|
using var content = new StringContent(payload.ToString(Formatting.None), Encoding.UTF8, "application/json");
|
|
|
|
using var resp = await httpClient.PostAsync(url, content).ConfigureAwait(false);
|
|
resp.EnsureSuccessStatusCode();
|
|
|
|
var json = await resp.Content.ReadAsStringAsync().ConfigureAwait(false);
|
|
return JObject.Parse(json);
|
|
}
|
|
|
|
|
|
private static void RequireFields(JObject info, string nameOrKey, params string[] fields)
|
|
{
|
|
foreach (var f in fields)
|
|
{
|
|
var v = info[f];
|
|
if (v == null || v.Type == JTokenType.Null)
|
|
throw new ArgumentException($"Model {nameOrKey} is missing required field: {f}");
|
|
}
|
|
}
|
|
|
|
private static long GenerateUInt32Seed()
|
|
{
|
|
Span<byte> bytes = stackalloc byte[4];
|
|
RandomNumberGenerator.Fill(bytes);
|
|
uint u = BitConverter.ToUInt32(bytes);
|
|
return (long)u; // preserve full 0..4294967295 range
|
|
}
|
|
|
|
|
|
private static JObject Edge(string srcNode, string srcField, string dstNode, string dstField) => new JObject
|
|
{
|
|
["source"] = new JObject { ["node_id"] = srcNode, ["field"] = srcField },
|
|
["destination"] = new JObject { ["node_id"] = dstNode, ["field"] = dstField }
|
|
};
|
|
|
|
|
|
|
|
private async Task<JObject> CreateText2ImgGraph(
|
|
string prompt,
|
|
string negativePrompt = "",
|
|
string modelKey = null,
|
|
string loraKey = null,
|
|
double loraWeight = 1.0,
|
|
string vaeKey = null,
|
|
int width = 512,
|
|
int height = 512,
|
|
int steps = 30,
|
|
double cfgScale = 7.5,
|
|
string scheduler = "euler",
|
|
long? seed = null)
|
|
{
|
|
// 1) Use default model if not specified: pick first "sd-1" from main list
|
|
if (string.IsNullOrEmpty(modelKey))
|
|
{
|
|
var models = await ListModels("main");
|
|
foreach (var token in models)
|
|
{
|
|
if (token is JObject m && string.Equals(m.Value<string>("base"), "sd-1", StringComparison.OrdinalIgnoreCase))
|
|
{
|
|
modelKey = m.Value<string>("key");
|
|
break;
|
|
}
|
|
}
|
|
|
|
if (string.IsNullOrEmpty(modelKey))
|
|
throw new ArgumentException("No suitable model found (sd-1)", nameof(modelKey));
|
|
}
|
|
|
|
// 2) Get model information
|
|
var modelInfo = await GetModelInfo(modelKey);
|
|
if (modelInfo == null)
|
|
throw new ArgumentException($"Model {modelKey} not found", nameof(modelKey));
|
|
if (modelInfo.Type != JTokenType.Object)
|
|
throw new ArgumentException($"Model {modelKey} returned invalid data type: {modelInfo.Type}", nameof(modelKey));
|
|
|
|
// 3) Validate required fields
|
|
RequireFields(modelInfo, modelKey, "key", "hash", "name", "base", "type");
|
|
|
|
// 4) Generate random 32-bit seed if not provided (0..2^32-1)
|
|
if (seed == null)
|
|
seed = GenerateUInt32Seed();
|
|
|
|
// 5) Detect SDXL
|
|
bool isSdxl = string.Equals(modelInfo.Value<string>("base"), "sdxl", StringComparison.OrdinalIgnoreCase);
|
|
|
|
// 6) Build nodes
|
|
var nodes = new JObject
|
|
{
|
|
// Main model loader
|
|
["model_loader"] = new JObject
|
|
{
|
|
["type"] = isSdxl ? "sdxl_model_loader" : "main_model_loader",
|
|
["id"] = "model_loader",
|
|
["model"] = new JObject
|
|
{
|
|
["key"] = modelInfo.Value<string>("key"),
|
|
["hash"] = modelInfo.Value<string>("hash"),
|
|
["name"] = modelInfo.Value<string>("name"),
|
|
["base"] = modelInfo.Value<string>("base"),
|
|
["type"] = modelInfo.Value<string>("type")
|
|
}
|
|
},
|
|
|
|
// Positive prompt
|
|
["positive_prompt"] = new JObject
|
|
{
|
|
["type"] = isSdxl ? "sdxl_compel_prompt" : "compel",
|
|
["id"] = "positive_prompt",
|
|
["prompt"] = prompt
|
|
},
|
|
|
|
// Negative prompt
|
|
["negative_prompt"] = new JObject
|
|
{
|
|
["type"] = isSdxl ? "sdxl_compel_prompt" : "compel",
|
|
["id"] = "negative_prompt",
|
|
["prompt"] = negativePrompt
|
|
},
|
|
|
|
// Noise generation
|
|
["noise"] = new JObject
|
|
{
|
|
["type"] = "noise",
|
|
["id"] = "noise",
|
|
["seed"] = seed,
|
|
["width"] = width,
|
|
["height"] = height,
|
|
["use_cpu"] = false
|
|
},
|
|
|
|
// Denoise latents
|
|
["denoise"] = new JObject
|
|
{
|
|
["type"] = "denoise_latents",
|
|
["id"] = "denoise",
|
|
["steps"] = steps,
|
|
["cfg_scale"] = cfgScale,
|
|
["scheduler"] = scheduler,
|
|
["denoising_start"] = 0,
|
|
["denoising_end"] = 1
|
|
},
|
|
|
|
// Latents to image
|
|
["latents_to_image"] = new JObject
|
|
{
|
|
["type"] = "l2i",
|
|
["id"] = "latents_to_image"
|
|
},
|
|
|
|
// Save image
|
|
["save_image"] = new JObject
|
|
{
|
|
["type"] = "save_image",
|
|
["id"] = "save_image",
|
|
["is_intermediate"] = false
|
|
}
|
|
};
|
|
|
|
// SDXL: add style fields (matches your Python **kwargs expansions)
|
|
if (isSdxl)
|
|
{
|
|
(nodes["positive_prompt"] as JObject)["style"] = prompt;
|
|
(nodes["negative_prompt"] as JObject)["style"] = "";
|
|
}
|
|
|
|
// 7) Optional: LoRA
|
|
if (!string.IsNullOrEmpty(loraKey))
|
|
{
|
|
var loraInfo = await GetModelInfo(loraKey);
|
|
if (loraInfo == null)
|
|
throw new ArgumentException($"LoRA model {loraKey} not found", nameof(loraKey));
|
|
RequireFields(loraInfo, loraKey, "key", "hash", "name", "base", "type");
|
|
|
|
nodes["lora_loader"] = new JObject
|
|
{
|
|
["type"] = "lora_loader",
|
|
["id"] = "lora_loader",
|
|
["lora"] = new JObject
|
|
{
|
|
["key"] = loraInfo.Value<string>("key"),
|
|
["hash"] = loraInfo.Value<string>("hash"),
|
|
["name"] = loraInfo.Value<string>("name"),
|
|
["base"] = loraInfo.Value<string>("base"),
|
|
["type"] = loraInfo.Value<string>("type")
|
|
},
|
|
["weight"] = loraWeight
|
|
};
|
|
}
|
|
|
|
// 8) Optional: VAE override
|
|
if (!string.IsNullOrEmpty(vaeKey))
|
|
{
|
|
var vaeInfo = await GetModelInfo(vaeKey);
|
|
if (vaeInfo == null)
|
|
throw new ArgumentException($"VAE model {vaeKey} not found", nameof(vaeKey));
|
|
RequireFields(vaeInfo, vaeKey, "key", "hash", "name", "base", "type");
|
|
|
|
nodes["vae_loader"] = new JObject
|
|
{
|
|
["type"] = "vae_loader",
|
|
["id"] = "vae_loader",
|
|
["vae_model"] = new JObject
|
|
{
|
|
["key"] = vaeInfo.Value<string>("key"),
|
|
["hash"] = vaeInfo.Value<string>("hash"),
|
|
["name"] = vaeInfo.Value<string>("name"),
|
|
["base"] = vaeInfo.Value<string>("base"),
|
|
["type"] = vaeInfo.Value<string>("type")
|
|
}
|
|
};
|
|
}
|
|
|
|
|
|
var edges = new JArray();
|
|
|
|
// Determine sources
|
|
bool hasLora = !string.IsNullOrEmpty(loraKey);
|
|
string unetSource = hasLora ? "lora_loader" : "model_loader";
|
|
string clipSource = hasLora ? "lora_loader" : "model_loader";
|
|
string vaeSource = !string.IsNullOrEmpty(vaeKey) ? "vae_loader" : "model_loader";
|
|
|
|
// If using LoRA, connect model_loader -> lora_loader (unet & clip)
|
|
if (hasLora)
|
|
{
|
|
edges.Add(Edge("model_loader", "unet", "lora_loader", "unet"));
|
|
edges.Add(Edge("model_loader", "clip", "lora_loader", "clip"));
|
|
// Note: lora_loader doesn't have clip2; SDXL clip2 comes from model_loader directly (handled below)
|
|
}
|
|
|
|
// Connect UNet to denoise
|
|
edges.Add(Edge(unetSource, "unet", "denoise", "unet"));
|
|
|
|
// Connect CLIP to prompts
|
|
edges.Add(Edge(clipSource, "clip", "positive_prompt", "clip"));
|
|
edges.Add(Edge(clipSource, "clip", "negative_prompt", "clip"));
|
|
|
|
// SDXL: connect clip2 from model_loader to both prompts
|
|
if (isSdxl)
|
|
{
|
|
edges.Add(Edge("model_loader", "clip2", "positive_prompt", "clip2"));
|
|
edges.Add(Edge("model_loader", "clip2", "negative_prompt", "clip2"));
|
|
}
|
|
|
|
// Prompts -> denoise conditioning
|
|
edges.Add(Edge("positive_prompt", "conditioning", "denoise", "positive_conditioning"));
|
|
edges.Add(Edge("negative_prompt", "conditioning", "denoise", "negative_conditioning"));
|
|
|
|
// Noise -> denoise
|
|
edges.Add(Edge("noise", "noise", "denoise", "noise"));
|
|
|
|
// Denoise -> l2i, and VAE -> l2i
|
|
edges.Add(Edge("denoise", "latents", "latents_to_image", "latents"));
|
|
edges.Add(Edge(vaeSource, "vae", "latents_to_image", "vae"));
|
|
|
|
// l2i -> save_image
|
|
edges.Add(Edge("latents_to_image", "image", "save_image", "image"));
|
|
|
|
// 7) Return final graph object
|
|
var graph = new JObject
|
|
{
|
|
["id"] = "text2img_graph",
|
|
["nodes"] = nodes,
|
|
["edges"] = edges
|
|
};
|
|
|
|
return graph;
|
|
}
|
|
|
|
|
|
private async Task<string> GenerateImageUrl(JObject arguments)
|
|
{
|
|
if (arguments == null) throw new ArgumentNullException(nameof(arguments));
|
|
|
|
// --- Extract parameters (with defaults) ---
|
|
string prompt = arguments.Value<string>("prompt")
|
|
?? throw new ArgumentException("Argument 'prompt' is required.", nameof(arguments));
|
|
string negativePrompt = arguments.Value<string>("negative_prompt") ?? "";
|
|
int width = arguments.Value<int?>("width") ?? 512;
|
|
int height = arguments.Value<int?>("height") ?? 512;
|
|
int steps = arguments.Value<int?>("steps") ?? 30;
|
|
double cfgScale = arguments.Value<double?>("cfg_scale") ?? 7.5;
|
|
string scheduler = arguments.Value<string>("scheduler") ?? "euler";
|
|
long? seed = arguments.Value<long?>("seed");
|
|
string modelKey = arguments.Value<string>("model_key");
|
|
string loraKey = arguments.Value<string>("lora_key");
|
|
double loraWeight = arguments.Value<double?>("lora_weight") ?? 1.0;
|
|
string vaeKey = arguments.Value<string>("vae_key");
|
|
|
|
|
|
// --- Create graph ---
|
|
JObject graph = await CreateText2ImgGraph(
|
|
prompt: prompt,
|
|
negativePrompt: negativePrompt,
|
|
modelKey: modelKey,
|
|
loraKey: loraKey,
|
|
loraWeight: loraWeight,
|
|
vaeKey: vaeKey,
|
|
width: width,
|
|
height: height,
|
|
steps: steps,
|
|
cfgScale: cfgScale,
|
|
scheduler: scheduler,
|
|
seed: seed
|
|
);
|
|
|
|
// --- Enqueue ---
|
|
JObject enqueueResult = await EnqueueGraph(graph);
|
|
string batchId = enqueueResult.SelectToken("batch.batch_id")?.Value<string>();
|
|
if (string.IsNullOrEmpty(batchId))
|
|
throw new InvalidOperationException("Enqueue response did not contain 'batch.batch_id'.");
|
|
|
|
UnityEngine.Debug.Log($"Enqueued batch {batchId}, waiting for completion...");
|
|
|
|
// --- Wait for completion ---
|
|
JObject completed = await WaitForCompletion(batchId);
|
|
|
|
// --- Extract image output ---
|
|
var outputs = completed.SelectToken("result.outputs") as JObject;
|
|
if (outputs != null)
|
|
{
|
|
foreach (var prop in outputs.Properties())
|
|
{
|
|
var output = prop.Value as JObject;
|
|
if (output?.Value<string>("type") == "image_output")
|
|
{
|
|
string imageName = output.SelectToken("image.image_name")?.Value<string>();
|
|
if (string.IsNullOrEmpty(imageName))
|
|
continue;
|
|
|
|
// Resolve relative URL for the image (API-dependent)
|
|
string imageRelativeUrl = await GetImageUrl(imageName);
|
|
return imageRelativeUrl;
|
|
}
|
|
}
|
|
}
|
|
|
|
throw new InvalidOperationException("Failed to generate image: no image_output found in result.");
|
|
}
|
|
|
|
public async Task<byte[]> GenerateImage(string prompt)
|
|
{
|
|
JObject args = new JObject()
|
|
{
|
|
["prompt"] = prompt,
|
|
["width"] = 512,
|
|
["height"] = 512,
|
|
["model_key"] = MODEL_KEY,
|
|
};
|
|
|
|
string imageUrl = await GenerateImageUrl(args);
|
|
|
|
|
|
var req = new HttpRequestMessage(HttpMethod.Get, imageUrl);
|
|
using var resp = await httpClient.SendAsync(req, HttpCompletionOption.ResponseHeadersRead);
|
|
resp.EnsureSuccessStatusCode();
|
|
|
|
return await resp.Content.ReadAsByteArrayAsync();
|
|
}
|
|
}
|