Default PaintPlus OpenAI provider to gpt-image-2, fix GPU quick-start

dall-e-2/dall-e-3 retired May 12 2026 and gpt-image-1 deprecates Oct 23
2026, so move every OpenAI default (config.py, both provider classes,
both compose files, .env.example, the in-app provider-settings dropdown,
README) to gpt-image-2 for both generation and edits. Also fix response
parsing in ai_provider.py's OpenAIProvider, which never sent a model
param and assumed a url response — gpt-image-1/2 only return b64_json.

Separately, .env.example shipped AI_PROVIDER=replicate by default, but
replicate has no driver in remote_provider.py, so following the
documented "cp .env.example .env" setup silently broke every AI call
and defeated the GPU quick-start (an explicit non-empty .env value
overrides docker-compose.gpu.yml's own local_gpu fallback). Default to
local_gpu instead, mark replicate/stability as not-yet-implemented, and
recommend Lykon/dreamshaper-8-inpainting as a hands/face-tuned
HF_MODEL_INPAINT override for 4-6GB cards (Quadro P2200, GTX 1060/1660).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Nb2vJ8W7bHKx1JXVvpCraH

paintplus/.env.example

@@ -19,18 +19,22 @@
 #   local_gpu = FREE, runs on YOUR GPU — best option if you have an NVIDIA card
 #               (use docker-compose.gpu.yml — models auto-download on first use)
 #   mock      = Free, returns original image unchanged (UI testing only)
-#   openai    = DALL-E 3 / gpt-image-1 (~$0.02-0.04/image)
-#   stability = Stability AI SDXL (~$0.01/image)
-#   replicate = Multiple models (~$0.002-0.03/image)
+#   openai    = gpt-image-2 generation + edits (~$0.006-0.21/image, see below)
+#   stability = NOT IMPLEMENTED YET — config field exists but no driver in
+#               remote_provider.py; setting this breaks every AI call
+#   replicate = NOT IMPLEMENTED YET — same as above (Replicate IS used for
+#               Smart Select's SAM fallback, but that's a separate code path)
 #   invokeai  = Self-hosted InvokeAI running on another machine
 #   comfyui   = Self-hosted ComfyUI running on another machine
 #
 # GPU QUICK-START:
 #   docker compose -f docker-compose.gpu.yml up --build
-#   (AI_PROVIDER defaults to local_gpu in that compose file)
+#   (AI_PROVIDER defaults to local_gpu in that compose file — but only if this
+#   var is unset/blank; since AI_PROVIDER=local_gpu is set explicitly below,
+#   that's what you get either way)
 # =============================================================================
 
-AI_PROVIDER=replicate
+AI_PROVIDER=local_gpu
 
 # ── Local GPU settings (only relevant when AI_PROVIDER=local_gpu) ────────────
 # Auto-download HuggingFace models on first request (true/false)
@@ -40,7 +44,11 @@ LOCAL_GPU_MAX_PIPELINES=2
 # HuggingFace token — only needed for gated/private models
 #HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
 # Override auto-selected model for any operation (leave blank = auto by VRAM tier)
-#HF_MODEL_INPAINT=your-org/your-inpaint-model
+# On 4-6GB cards (Quadro P2200, GTX 1060/1660, etc.) the auto-tier picks either
+# slow SDXL+CPU-offload or a generic (not hand/face-tuned) SD checkpoint.
+# Lykon/dreamshaper-8-inpainting is SD1.5-based (1.7GB, fast, no offload needed)
+# and noticeably better on hands/faces — worth forcing on small cards:
+#HF_MODEL_INPAINT=Lykon/dreamshaper-8-inpainting
 #HF_MODEL_TXT2IMG=your-org/your-txt2img-model
 #HF_MODEL_IMG2IMG=your-org/your-img2img-model
 # ─────────────────────────────────────────────────────────────────────────────
@@ -76,17 +84,19 @@ LOCAL_GPU_MAX_PIPELINES=2
 REPLICATE_API_KEY=r8_PASTE_YOUR_KEY_HERE
 
 # ───────────────────────────────────────────────────────────────────────────
-# OPENAI (cloud, dall-e-3 / gpt-image-1)
+# OPENAI (cloud, gpt-image-2)
 # Get key at: https://platform.openai.com/api-keys
 # AI_PROVIDER=openai
 # ───────────────────────────────────────────────────────────────────────────
+# dall-e-2 and dall-e-3 were retired May 12, 2026; gpt-image-1 deprecates
+# Oct 23, 2026. gpt-image-2 is the current model and handles both
+# generation and masked edits (inpaint/img2img/outpaint) — no org
+# verification step needed, unlike gpt-image-1.
+# Pricing (1024x1024): ~$0.006 low / $0.053 medium / $0.211 high quality.
 #OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
-#OPENAI_MODEL=dall-e-3
+#OPENAI_MODEL=gpt-image-2
 # Model for inpaint/img2img/outpaint (the /v1/images/edits endpoint).
-# dall-e-3 does NOT support edits at all — keep this on gpt-image-1 for
-# ChatGPT-comparable edit quality, or set dall-e-2 for the older/cheaper option.
-# Note: gpt-image-1 may require completing org verification at platform.openai.com.
-#OPENAI_EDIT_MODEL=gpt-image-1
+#OPENAI_EDIT_MODEL=gpt-image-2
 
 # ───────────────────────────────────────────────────────────────────────────
 # INVOKEAI (self-hosted, best for Flux/SDXL)