sqlite-web 0.7.2
gitea.db
action
Create
Query
access
access_token
action
action_artifact
action_run
action_run_index
action_run_job
action_runner
action_runner_token
action_schedule
action_schedule_spec
action_task
action_task_output
action_task_step
action_tasks_version
action_variable
app_state
attachment
auth_token
badge
branch
collaboration
comment
commit_status
commit_status_index
commit_status_summary
commit_sync_log
commit_sync_status
dbfs_data
dbfs_meta
deploy_key
email_address
email_hash
external_login_user
follow
gpg_key
gpg_key_import
hook_task
issue
issue_assignees
issue_content_history
issue_dependency
issue_index
issue_label
issue_pin
issue_user
issue_watch
label
language_stat
lfs_lock
lfs_meta_object
login_source
milestone
mirror
notice
notification
oauth2_application
oauth2_authorization_code
oauth2_grant
org_user
package
package_blob
package_blob_upload
package_cleanup_rule
package_file
package_property
package_version
project
project_board
project_issue
protected_branch
protected_tag
public_key
pull_auto_merge
pull_request
push_mirror
reaction
release
renamed_branch
repo_archiver
repo_hidden_file
repo_indexer_status
repo_license
repo_redirect
repo_topic
repo_transfer
repo_unit
repository
review
review_state
secret
session
sqlite_sequence
star
stopwatch
system_setting
task
team
team_invite
team_repo
team_unit
team_user
topic
tracked_time
two_factor
upload
user
user_badge
user_blocking
user_open_id
user_redirect
user_setting
version
watch
webauthn_credential
webhook
Toggle helper tables
Structure
Content
Query
Insert
Drop
Import
Export
Update row 16221 in action
id
Primary key.
INTEGER NOT NULL
user_id
INTEGER
op_type
INTEGER
act_user_id
INTEGER
repo_id
INTEGER
comment_id
INTEGER
is_deleted
INTEGER NOT NULL (default 0
ref_name
refs/heads/feat/qwen-image-inference
TEXT
is_private
INTEGER NOT NULL (default 0
content
{"Commits":[{"Sha1":"ac5f340c0422af2a5c3122289142e8517dc434c5","Message":"fix(sidecar): Qwen-Image-Edit 未指定尺寸时保留底图原生分辨率\n\nQwenImageEditPlus 默认把输入塌到 ~1MP 规范面积,密集中文标题在 1MP + 低步数下\n笔画糊/字形崩。新增 _fit_edit_dims:未显式传 width/height 时按底图原生分辨率\n(保纵横比、向下取整 16 倍数、封顶 QWEN_EDIT_MAX_AREA=1.5MP 兜内存/耗时)驱动,\n调用方仍可显式覆盖。\n\n实测 ktv-poster-master(1672x941) 改字:4 步@1.06MP 文字 melted → 20 步@1.49MP\n(1632x912)文字笔画完整清晰、字形正确,无高分辨率伪影。清晰度决定因素是步数\n(生产默认 40),分辨率为次因;两者叠加达到可用海报质量。\n\nCo-Authored-By: Claude Opus 4.8 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luojuhai@luojuhaideMacBook-Pro.local","AuthorName":"luojuhai","CommitterEmail":"luojuhai@luojuhaideMacBook-Pro.local","CommitterName":"luojuhai","Timestamp":"2026-07-07T07:39:05+08:00"},{"Sha1":"973dde30db19928a0937f02a102979d473cce644","Message":"feat(material-factory): 图像模型真实推理对接 + 双后端渲染管线闭环\n\n把嗨设工坊从「权重就绪、推理是桩」推进到「本机真跑转绿」:\n\nsidecar (services/model-sidecar)\n- /v1/generate 真调 Z-Image-Turbo / Qwen-Image-2512(modelKey 路由)\n- /v1/edit 真调 Qwen-Image-Edit-2511(QwenImageEditPlusPipeline)\n- capability() 依赖齐全时开启 qwen-image-master/edit(128 GiB 主机 loadable)\n- 单 pipeline 常驻、kind 切换释放旧模型,防双 20B 同驻 OOM\n- 修 torchvision 隐藏依赖:Qwen2VLProcessor→Qwen2VLVideoProcessor 硬依赖它,\n 缺失则 from_pretrained ImportError(进 pyproject zimage extra)\n- 加载失败包 try/except → 结构化 503(不再冒 500)\n- SIDECAR_QWEN_STEPS:生产默认 40 步,CI/验收可降步提速(真引擎真像素)\n\n双后端 (NestJS + Fastify 同契约)\n- render-pipeline 按 kind 路由:APPEARANCE/SEMANTIC_EDIT→sidecarEdit,\n DESIGN_MASTER/TURBO_VARIANT→sidecarGenerate;job 状态机真流转 + 乐观锁 + outbox 同 tx\n- 修 undici 300s 传输超时:Node 内置 fetch 的 headersTimeout/bodyTimeout 默认 300s\n 会把分钟级同步推理在 ~302s 中止成假 SidecarUnavailableError(AbortSignal 压不住)。\n 改用 undici@6 request() + 请求级 headersTimeout/bodyTimeout 对齐 timeoutMs\n\n验收\n- DATABASE_URL=... INFERENCE_RUN_QWEN=1 pnpm check:inference 双后端各 6/6 绿\n- Qwen-Image-Edit 真改图 SUCCEEDED(金色中文标题精确改写 + OCR gate 联动)\n- pnpm check 全部静态门禁绿;CLAUDE.md(=AGENTS.md) 回灌 G14/真跑证据\n\n仍打开(诚实缺口):render-pipeline 同步端点(G13)、异步 worker 化、\n单 MPS 多进程串行化、Qwen nightly/self-hosted 验收。\n\nCo-Authored-By: Claude Opus 4.8 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luojuhai@luojuhaideMacBook-Pro.local","AuthorName":"luojuhai","CommitterEmail":"luojuhai@luojuhaideMacBook-Pro.local","CommitterName":"luojuhai","Timestamp":"2026-07-07T06:35:04+08:00"}],"HeadCommit":{"Sha1":"ac5f340c0422af2a5c3122289142e8517dc434c5","Message":"fix(sidecar): Qwen-Image-Edit 未指定尺寸时保留底图原生分辨率\n\nQwenImageEditPlus 默认把输入塌到 ~1MP 规范面积,密集中文标题在 1MP + 低步数下\n笔画糊/字形崩。新增 _fit_edit_dims:未显式传 width/height 时按底图原生分辨率\n(保纵横比、向下取整 16 倍数、封顶 QWEN_EDIT_MAX_AREA=1.5MP 兜内存/耗时)驱动,\n调用方仍可显式覆盖。\n\n实测 ktv-poster-master(1672x941) 改字:4 步@1.06MP 文字 melted → 20 步@1.49MP\n(1632x912)文字笔画完整清晰、字形正确,无高分辨率伪影。清晰度决定因素是步数\n(生产默认 40),分辨率为次因;两者叠加达到可用海报质量。\n\nCo-Authored-By: Claude Opus 4.8 \u003cnoreply@anthropic.com\u003e\n","AuthorEmail":"luojuhai@luojuhaideMacBook-Pro.local","AuthorName":"luojuhai","CommitterEmail":"luojuhai@luojuhaideMacBook-Pro.local","CommitterName":"luojuhai","Timestamp":"2026-07-07T07:39:05+08:00"},"CompareURL":"luoanwu/image-generation/compare/fc2ce503c94d763b1d077e7c734720a0c58e23a8...ac5f340c0422af2a5c3122289142e8517dc434c5","Len":2}
TEXT
created_unix
INTEGER
Update
Cancel